Continual Learning in Medical Imaging from Theory to Practice: A Survey and Practical Analysis (2024)

1 Introduction

Deep Learning (DL) algorithms are rapidly gaining relevance in medical imaging, enabling segmentation (Chen etal., 2020; Hofmanninger etal., 2020b), classification (Kumar etal., 2023), and detection Halder etal. (2020) of anatomical structures and anomalies (Schlegl etal., 2019; Qazi etal., 2023) relevant for diagnosis, prediction, or prognosis. In some cases, their capabilities surpass those of human experts (Walsh etal., 2018; DeFauw etal., 2018), making them a central tool in the advancement of the use of medical data to support clinical decisions (Ahsan etal., 2022; Adlung etal., 2021). Although there are several DL works in medicine, the task is often oversimplified by assuming a static, centralized data source and neglecting the dynamic nature of the evolving data.

In general, a model is trained and tested on different splits of the same dataset, and the models are expected to generalize well to other datasets of the same nature. However, this does not reflect the reality of real-world deployment because of frequent changes in data distribution across various institutions, diagnostic devices, and population demographics (Cohen etal., 2021). For Instance, in 2020, Google developed a DL system trained on a vast heterogeneous dataset that achieved a high performance in an offline setting for detecting diabetic eye diseases (Beede etal., 2020). However, when this system is deployed, its performance drops drastically, highlighting the severity of the problem and emphasizing the need for robust systems. These issues can be partially addressed through domain adaptation and generalization.

In real-world scenarios, a model trained at a specific point in time struggles to maintain optimal performance owing to the continuous evolution of the data distributions. Retraining the model on the entire dataset is impractical because substantial computational resources are required, and privacy concerns often restrict access to previous data, exacerbating this issue. Additionally, sequential fine-tuning of models can lead to overfitting of the latest task, resulting in a decline in the performance on previous data.

Continual learning (CL) has emerged as a promising approach for training deep neural networks in order to overcome these challenges. CL offers a solution that enables neural networks to learn continuously, resulting in a more resilient model capable of accommodating new tasks without forgetting previous tasks. Although the model remains static when deployed in a static environment, it continuously updates itself in a CL setting, thereby yielding a more robust and versatile model. Addressing the challenges mentioned above requires consideration of the rigidity and plasticity of the neural networks. Rigidity, the ability of a neural network to preserve previous knowledge, and plasticity, the ability to learn new tasks are crucial factors. A model with high plasticity can learn new tasks over time, achieving good performance on future tasks but deteriorating performance on older tasks, a phenomenon known as "catastrophic forgetting". An ideal neural network should balance rigidity and plasticity to demonstrate its ability to learn new tasks without forgetting prior knowledge. CL has emerged as a promising approach for balancing these properties, enabling neural networks to learn continuously and adapt to evolving data distribution.

1.3 Paper Organisation

The rest of this paper is organized as follows. We start by mathematically formalizing CL in Section 2. Then, we discuss and formulate the three types of CL: Task Incremental Learning (Section 2.1), Domain Incremental Learning (Section 2.2), and Class Incremental Learning (Section 2.3). Within this section, we first discuss the theory of these three types of CL before mathematically formulating them with an example. In Section 3, we review the recent advances in CL methods in the medical field and divide the studies on the approach they took to alleviate forgetting. Then, in Section 4, we discussed the overview of the studies along with the current research trend (section 4.1) and the practicality of these studies (section 4.2). Section 5 provides insight into future directions in the field and section 6 concludes the paper.

2 Background

In this section, we present a comprehensive conceptual framework that includes the mathematical formulation of three primary types of CL: task, domain, and class incremental learning. The framework highlights the potential of these three types and how they can be applied to various advanced medical tasks. The purpose of this framework is to underline the essential concepts of CL, and how diagnostic and prognostic models with CL capabilities are robust and adaptable. The framework showcases the applications of CL in the medical field through Figure 3, thereby highlighting CL as a foundational feature in models that need to incrementally learn from new data sources while retaining previously learned knowledge.

To formalize the CL framework, we consider a model M represented by a parametric function f, which aims to learn the mapping between the input space X and the output space Y. Here, X represents the medical data, including sensor readings, images, and electronic health records (EHR), while Y includes diagnostic classifications, tumor or organ segmentation, and other medical-related predictions. Function f is parameterized by weights $\theta$. In a CL setting, f must adapt these parameters as new data becomes available, such as tasks, domains, and/or classes.

One of the biggest challenges in CL is the phenomenon of catastrophic forgetting, in which learning new information leads to forgetting previously known knowledge. The main objective is to optimize the parameters $\theta$ of the function f over a data distribution D, which changes over time, reflecting new tasks (T), domains (D), or classes (C).

where t represents the time, X_t and Y_t are the input and output at time t, the loss function $\mathcal{L}$ calculates the difference between the model prediction and the true outputs, $\Omega$ penalizes the change in parameters important for the previously learned knowledge, and $\lambda$ is a hyperparameter that balances the retention of old knowledge and new learning.

In the remainder of this section, we use this mathematical framework to formalize the three types of CL techniques. Each presents its own challenges and requires a tailored approach to alleviate forgetting and seamlessly learning new knowledge. This shared vocabulary based on mathematical formulation will help communicate solutions for CL problems.

2.1 Task-Incremental Learning

Theory: In the scope of CL, task-incremental learning refers to the model’s ability to learn new tasks sequentially, using the previously learned knowledge. This method emphasizes on clear separation of tasks with distinct objectives. It is always explicitly defined for the model what task to perform by presenting a task identifier, thereby avoiding any interference between tasks. This approach allows the model to incrementally learn new tasks without forgetting the old ones.

Formulation: Formally, given a non-stationary n tasks ${T_{1},T_{2},T_{3},...,T_{n}}$, that require to learn the mapping from $X_{i}$ to $Y_{i}$ for each task, the task-incremental protocol relies on the task identifier $t$ to distinguish between the tasks, and thereby, performing the tasks individually. In this design, the model can be developed to include, for example, an individual output layer for each task or include task-distinctive components in the network. Mathematically, this can be presented as optimizing the following objective function for each task:

$$\arg\min_{\theta_{\tau+1}}\left[\mathcal{L}\left(f\left(X_{T_{i}},\theta_{\tau$$

(2)

where $\theta_{\tau}$ are the parameters of the model-specific task $\tau$, $\mathcal{L}$ is the loss function and $\Omega$ is the regularization term that conserves the knowledge from previously learned tasks.

Example: Figure 3 (left) provides an illustration of task-incremental learning in medical field. Initially, the model is trained to segment brain tumors. As the new data arrives, the model learns to identify lung tumors, using a new set of parameters while retaining the knowledge from previously learned brain tumor segmentation tasks. This sequential learning process is represented in the same model by two separate pathways, each activated by its corresponding task identifier.

3 Taxonomy

Numerous studies have addressed incremental learning in medical images in recent years. We organize these methods taxonomically from the approach they take to implement, that is, regularization-based, replay-based, and dynamic model-based. Replay-based methods concentrate on solving the CL with exemplars, which can be further classified into data memory and feature replay. Regularization-based methods regularize the model parameters from drifting away. Finally, dynamic model methods either isolate parameters based on the task or expand the model structure to accommodate more task knowledge. The following sections discuss CL methods using these three approaches.

3.1 Regularization Approaches

Regularization-based continual learning offers an effective solution for learning new tasks while preventing catastrophic forgetting of previous ones. The key concept behind this approach is to minimize the drift in the learned feature or parameter space of the model, as shown in Figure 4. By doing so, the model can learn new tasks without inducing significant changes in the parameters that are useful for old tasks, enabling it to retain prior knowledge (Wang etal., 2023). Regularization-based CL incorporates additional loss terms to prevent model drift between old and new tasks. This is typically achieved by maintaining a copy of the model from the previous task or by preserving the model predictions. In medical applications, regularization-based CL are valuable because they do not require access to stored samples, which is a challenge because of privacy concerns. In addition, it does not dynamically increase network size. This makes regularization-based CL models lightweight and versatile, making them particularly suitable for various medical settings. Regularization-based methods fall into two primary categories: parameter regularization and functional regularization, depending on whether they focus on the network parameters or output.

3.1.1 Parameter Regularization

Parameter Regularization focuses on the change in network parameters when learning new tasks. This is achieved by introducing an additional loss term to penalize alterations in the parameters deemed important for the previous tasks. Typically, this involves maintaining a frozen copy of the old model to calculate the change in network parameters with the introduction of new tasks. Several studies have demonstrated the effectiveness of this approach in the medical field. Baweja etal. (2018) incorporated a quadratic penalty in the loss function to address parameter changes important for prior tasks. It follows the Elastic Weight Consolidation (EWC) method using the Fisher information matrix to quantify each parameter’s contribution (Kirkpatrick etal., 2017). Their work comprehensively analyzed the effectiveness of the EWC method in medical imaging applications. Similarly, van Garderen etal. (2019) applied a strategy of penalizing changes in important parameters for brain MRI segmentation, demonstrating its efficacy in transfer learning by pre-training on a high-quality dataset before fine-tuning it on a smaller, lower-quality dataset. This provides an insight into the application of regularization-based CL methods in transfer learning for low-quality datasets, which are common in the medical domain. Another study by Lenga etal. (2020) used EWC and Learning Without Forgetting (LWF) methods for multidomain chest X-ray classification and compared their performance against joint training. Their work showed that the LWF outperformed the EWC approach and achieved a backward transfer on par with joint training on the new domain conducted with 60-80% of the data reply of the previous domain. Chen and Tang (2022) combined the hybrid-averaging operation and EWC into CNN models for breast cancer image classification tasks and found that the ResNet (He etal., 2016) outperforms AlexNet (Krizhevsky etal., 2012) and DenseNet (Huang etal., 2017) in terms of average accuracy and average forgetting. In addition, Derakhshani etal. (2022) and Quarta etal. (2022) offered a benchmark for various regularization-based CL methods, including EWC (Kirkpatrick etal., 2017), LwF (Li and Hoiem, 2017), iCaRL (Rebuffi etal., 2017), MAS (Aljundi etal., 2018), and EEIL (Castro etal., 2018), and showed that iCaRL and EEIL provide superior performance compared to other methods.

More recently, Özgün etal. (2020) proposed a modified version of the memory-aware synapses (MAS) (Aljundi etal., 2018) method by imposing learning rate (LR) constraints on each parameter, instead of direct parameter regularization. In this approach, each parameter or an entire CNN kernel is assigned a specific LR based on its importance for the previous tasks. The authors also proposed freezing the important parameters and fine-tuning the unimportant parameters using LR constraints. Furthermore, Chen etal. (2021) introduced the Targeted Gradient Descent (TGD) method to fine-tune a model without forgetting previous knowledge by leveraging the Kernel Sparsity and Entropy (KSE) metric in a positron emission tomography (PET) reconstruction study. TGD provides an effective method of reusing redundant kernels in a pre-trained CNN for learning new tasks while preserving the balance between the rigidity and plasticity of the models. Another study by Bayasi etal. (2021) used a pruning criterion that allows a fixed network to learn new data domains sequentially over time by detecting the culprit units associated with the wrong classification. These network units are dedicated to the subsequent learning of new domains. MUSCLE (Liao etal., 2022) proposed the addition of model parameter constraints for pretraining a robust backbone in the classification, segmentation, and detection of X-ray images. It utilizes momentum updates between model parameters to prevent catastrophic forgetting during pretraining when moving from one task to another. Ranem etal. (2022) combined a ViT (Dosovitskiy etal., 2020) and nnUNet (Isensee etal., 2021) models for continual hippocampus segmentation using the EWC and PLOP methods(Douillard etal., 2021). Their proposed model utilizes the self-attention mechanism of transformers to maintain knowledge and consequently alleviate catastrophic forgetting, in contrast to the purely convolutional UNet. Nevertheless, interestingly, their experiments showed that applying regularization to the transformer component over time has an adverse effect on knowledge retention, as regularization disrupts the self-attention mechanism. In contrast, it is preferable to discourage substantial changes in the convolutional layers to retain prior knowledge without compromising the plasticity. Furthermore, Zhang etal. (2021, 2023a) proposed a comprehensive importance (CI) method for selecting important parameters based on their contribution to the shape and certainty of output segmentation. They demonstrated that maintaining such parameters can preserve the important shape and semantic information of the segmentation task in CL settings. In another study by Shu etal. (2022), regularization in the gradient space of weights was implemented. This involved updating the model weights using gradients orthogonal to the important parameters for the previous task(s). To further enhance the performance, they incorporated data replay into their methods. Furthermore, Wang and Luo (2022) introduced a technique to mitigate catastrophic forgetting in lung nodule detection studies. Their approach was based on a modified EWC and incorporated feature distillation for enhanced performance. Conversely, Liu etal. (2023) used a frozen copy of the model along with other techniques to adapt an off-the-shelf segmentation model for different anatomical regions. The two model copies were combined by taking a weighted average of the corresponding layers at the end of each training round, allowing for the retention of previous knowledge while acquiring new task-specific information.

3.1.2 Functional Regularization

Functional Regularization impedes the intermediate or final prediction of the model by using a distillation-based approach, where the old model acts as a teacher to the new model. This approach does not include any explicit constraints on the model parameters and only uses the model’s output to keep the learned feature space close between the tasks.

Kim etal. (2018) proposed learning an inverse function that maps the model’s logits to the feature space back and using it to regularize the feature space of the model in the next task. They demonstrated the effectiveness of their method in a multidomain, single-task X-ray classification scenario. An interesting domain incremental study provided a dual-distillation method combined with a replay to mitigate forgetting (Li etal., 2020). Dual-distillation is carried out using the logits from the old and a new fully fine-tuned model to train an updated model that performs well on new and old tasks. Recent work on manifold learning Akundi and Sivaswamy (2021) used a reformer network to augment any input domain image to lie in the same data manifold. This, combined with various regularization-based methods, was shown to be effective in reducing forgetting due to domain shifts in the clinical data. Moreover, a multimodal study by Patra etal. (2021) on ultrasound and eye-tracking data utilized distillation to train a model on a continual stream of data, whereas Patra etal. (2021) proposed an architecture that leverages the simultaneous availability of two or more datasets to learn a disentanglement between the content and domain in an adversarial manner. This allows a domain-invariant content representation that lays the basis for the continual semantic segmentation of MRI images.He etal. (2021) use a probability-map alignment scheme to integrate the previous and current map for a segmentation task. They further used knowledge distillation to transfer the knowledge from the integrated probability map to the current model, thereby improving the segmentation performance of the current model in the CL. Chakraborti etal. (2021) presents a semi-supervised continual learning method combined with a contrastive loss. Their approach was to first train an auto-encoder with a small amount of labeled data in the usual supervised manner, followed by an unsupervised adaptive regularization-based contrastive learning approach.

Another work by Liu etal. (2022b) introduced a method for multi-organ segmentation in CT images. Their methodology uses lightweight memory modules and a tailored regularization loss function, allowing the model to effectively utilize replayed examples to prevent forgetting. Here, the introduced loss function constrains the learned feature space to maintain the feature of an old class close to its mean representation within the memory module, while ensuring that the features of a new class are distinct from all representations of the old class. Furthermore, Tian etal. (2022) proposed a multi-scale multitask CL framework. Their method combined contrastive learning with distillation-based regularization and data rehearsal to enhance learning without catastrophic forgetting. Li etal. (2022a) used feature whitening to regularize the important domain in-variant features in a domain incremental cardiac image segmentation study. This feature regularization was augmented by a generative feature reply, where they proposed a base conditional generator along with a style module to produce images for replay. Furthermore, Roy etal. (2023) introduced the use of a mixed-curvature embedding space for knowledge distillation to prevent catastrophic forgetting. The authors employed a combination of Euclidean and hyperbolic spaces to project the learned feature embeddings, emphasizing their importance for complex data, especially in medical imaging. Another recent study by Gao etal. (2023) used a Kullback–Leibler (KL) divergence-based loss function to preserve the models’ knowledge by penalizing the change in output of the new and old models. More recently, the Fourier test–time adaptation (FTTA) method was presented for classifying medical images with domain gaps (Huang etal., 2023). This technique utilizes a Fourier-based domain adjustment strategy along with a multilevel consistency approach to regularize the model’s output. The proposed method is based on a multilevel consistency measurement designed for self-correcting predictions.

Overall, regularization-based methods offer an efficient and privately protected approach to address catastrophic forgetting in CL. These methods do not introduce any computational overhead or privacy concerns, making them well suited for real-life medical applications. However, standalone regularization-based CL methods often fall short of incrementally learning very different tasks, such as diverse class incremental learning, whereas their effectiveness becomes more pronounced in domain incremental settings. Recent attempts to combine regularization-based CL methods with other CL techniques have been beneficial in addressing this issue, which can be a promising direction for making regularization-based CL methods practical for medical applications.

3.2 Replay-Based Approaches

Replay-based methods in CL were designed to address the problem of catastrophic forgetting in sequential learning scenarios. These methods use a replay mechanism or memory buffer to preserve and reintroduce prior experiences while training new tasks to reduce forgetting, as depicted in Figure 5. This buffer typically contains samples, such as input-output pairs and representations of past data, preserving knowledge that the model has learned previously.

The ability of replay-based techniques to retain learned knowledge while adjusting to new information or tasks is a crucial feature that makes them highly advantageous and applicable to the medical domain. This ability to retain information is critical, as it ensures that the model retains the critical patterns, diagnoses, or treatment protocols that it has learned from previous experiences. Replay-based models facilitate the exchange of knowledge and skills among various healthcare facilities and regions. These strategies can also help with privacy and data constraints by effectively utilizing a smaller collection of experiences. This reduces the need to constantly access or store new patient data, while maintaining the confidentiality of sensitive information.

Venkataramani etal. (2019) use Memory-augmented neural networks(MANN) for continuous domain adaptation for the task of semantic segmentation. They proposed a single framework for domain adaptation in which images similar to the target image and their respective textual and shape features are fetched using wavelet features of the target image as a query. These context features, along with the target image features, were embedded into the model trained using only the source domain. Ravishankar etal. (2019) proposed a single approach for the different variants of lifelong learning, such as multi-task setting and single incremental task setting, aiming to ensure data privacy and mitigate catastrophic forgetting. They employed a pseudo-rehearsal strategy using a finite memory module and ensured class separation using a composite loss. Furthermore, they introduce feature transformers to learn new representations from the old data as well as the new data. Hofmanninger etal. (2020a) proposed a rehearsal method that infers the data shifts by adopting a dynamic memory module. Their memory update strategy was designed based on a fixed set of rules, and the features were replaced with a high-level metric calculated using gram matrices. Perkonigg etal. (2021) employs the same strategy but also introduces an additional module called the pseudo-domain (PD) model, which detects clusters of similar styles from the continuous stream of data. These pseudo-domains are seen as proxies for the unknown, real domains, and are used to balance the memory and training process (DM-PD). Srivastava etal. (2021) also proposed another domain incremental learning that leverages vector quantization approach under limited memory constraints to effectively store and replay hidden representations where no clear domain-shift boundaries exist and no a-priori knowledge is possible. Wang etal. (2021) introduced a method that integrates medical expertise into disease diagnosis by considering context and medical entity characteristics in order to transfer knowledge to new stages. Their work was among the pioneering studies exploring CL within medical domains, specifically focusing on diagnosing diseases based on clinical notes. Additionally, they introduced a new technique called embedding episodic memory and consolidation (E2MC) to avoid catastrophic forgetting in disease diagnosis tasks. Morgado etal. (2021) delve into continual learning within dermatology, employing four distinct methods: naive, elastic weight consolidation, averaged gradient episodic memory, and experience replay. These methodologies allow adaptation to new conditions while safeguarding previously acquired knowledge in the field.

To overcome the problem of insufficient sensitive patient data for digital biomarker model training owing to privacy laws, Shevchyk etal. (2022) proposed a Privacy-Preserving Synthetic Data Generation pipeline that uses a GAN to generate anonymous medical samples. Privacy is ensured using a Siamese Neural Network (SNN) with triplet loss to evaluate the generated data. The effectiveness of this method was demonstrated using a Respiratory Sound Classifier that collects respiratory symptom information and is relevant for conditions such as Asthma, COPD, or COVID-19.

Similarly, Liu etal. (2022a) pioneered a generative replay-based approach in semantic segmentation tasks for medical imaging by introducing SegGAN to prevent catastrophic forgetting. This GAN generates both previous images and their corresponding pixel-level labels, addressing privacy limitations while preserving the acquired knowledge for the subsequent learning stages. Additionally, they proposed a unique filtering mechanism that ensures the selection of high-quality generated data by maintaining consistency between the pseudo-labeling and generative replay techniques. Li etal. (2022c) propose the use of multiple generators to improve the quality of replayed samples. Instead of using generators to generate higher-resolution images, this study uses a fixed feature extractor and focuses on feature generation, which requires fewer parameters and fewer computations. Li etal. (2022b) introduced a fresh approach for CL in gesture recognition, specifically for adding new gesture classes. This method focuses on managing out-of-distribution (OOD) memory and improving rehearsal to consistently learn new gestures. Additionally, it incorporates an energy-based loss to penalize out-of-distribution gesture examples and further refine the model during training. Ayromlou etal. (2022) introduces an innovative data-free framework for incremental class learning. This approach initially creates synthetic data from the model trained on past classes to form a Class Impression, and then updates the model by merging the synthesized data with new class data. They employ various loss functions, such as cosine normalized cross-entropy, margin loss, and intra-domain contrastive loss, to tackle class imbalances, differentiate between previous and new classes, and ensure the model’s generalization from synthetic to real data.Zhang etal. (2022) introduce a new method, SMG (Synchronous Memorizability and Generalizability) -learning, to enhance both memorization and generalizability. It presents the Synchronous Gradient Alignment (SGA) objective of improving network memorization and adaptability across previous and new sites. To streamline optimization without excessive computational load, the dual-meta algorithm was developed. Moreover, to enhance rehearsal effectiveness, the replay buffer is modified to include diverse data from various sites, thus minimizing redundancy.

Bera etal. (2023) introduce a novel solution to improve image selection to create the memory bank by ranking and selecting images based on their contribution to the learning process. In situations such as medical applications where storing previous data is limited by privacy regulations, Chen etal. (2023) presented Generative Appearance Replay for continual Domain Adaptation (GarDA), an approach employing generative replay techniques. GarDA facilitates the gradual adjustment of a segmentation model to new domains by utilizing unlabeled data. Similarly, Thandiackal etal. (2023) proposed a generative feature-driven image replay with a dual-purpose discriminator to align features for continual learning in unlabeled domains, achieving state-of-the-art results in histopathology datasets. Unlike single-step unsupervised domain adaptation (UDA), continual adaptation across various domains enables the utilization and consolidation of knowledge from different domains throughout consecutive learning phases. Wei etal. (2023) proposes a data selection strategy using Variational autoencoders (VAE) and adversial network. Through this approach, they retain training efficiency by always selecting a fixed subset size from the entire dataset. Kim etal. (2024) proposes a continual learning for multicenter studies, eliminating the need for a central server and mitigating the risk of catastrophic forgetting of previously acquired knowledge. The framework involves a method selection process for continual learning, acknowledging that no single method suits all datasets and utilizes fake data generated by a generative adversarial network to evaluate the methods.

Replay-based techniques can be used in medical applications; however, doing so poses additional challenges with regard to model interpretability, regulatory framework validation, and ethical issues that must be addressed to ensure the safety and effectiveness of AI-driven healthcare solutions. Addressing these challenges is crucial for harnessing the full potential of CL techniques in the medical domain while ensuring patient safety and privacy.

3.3 Dynamic Model Approaches

Previous methods were either limited by the model’s parameter change in subsequent tasks or relied on rehearsal techniques. A more recent and promising solution involves intelligent introduction of new parameters for each task (Figure 6. In other words, shared features are maintained among tasks to avoid repeated training, similar to ensembling. Moreover, employing a set of new parameters enables more effective learning of task-specific features.

Dynamic models offer a direct solution by allocating specific segments of the model to distinct tasks. A study by Karani etal. (2018) addressed domain shifts in lifelong learning scenarios using domain-specific batch normalization layers for MRI segmentation. Variance in MRI image intensities, caused by multiple factors, leads to this domain shift (Preboske etal., 2006; Jovicich etal., 2006).

However, another challenge arises when we lack knowledge about the task or which sub-network to choose during inference for particular data. For example, if a model is trained on images from three different scanners, determining the domain during inference requires image metadata to activate the corresponding model segment. Gonzalez etal. (2020) tackled this issue by training separate UNet models for domain segmentation. To discern the domain during inference, they utilized separate Variational Autoencoders (VAEs) for each domain and selected the domain with the minimal VAE reconstruction loss. Another study by González etal. (2022) introduces a task-agnostic continual learning approach applicable across diverse deep learning architectures. Their method involved utilizing the Mahalanobis distance to identify domain changes. Upon detecting such changes, they expand the network, ensuring a nonlinear expansion. During inference, the network leverages this distance measurement to determine task ID, thereby facilitating efficient task identification. Similarly, Bayasi etal. (2023) selected a network based on a distance-based metric for each test dataset. In contrast, Zhao etal. (2023) used different prompts for each task and used a matching strategy to choose the correct prompt at inference.

Recently, advanced techniques have been applied to address the challenges of multiorgan segmentation. These methods operate in a class incremental setting, requiring the model to segment all the regions observed in previous tasks during inference. Zhang etal. (2023c) highlights the significance of dynamically expanding models to encompass new data and classes for segmenting multiple organs and tumors. Their architecture features lightweight class-specific heads that replace traditional output layers, enabling separate predictions for both new and previously learned classes. Leveraging extensive image-text co-training, the authors embedded contrastive language-image pretraining (CLIP) embeddings into organ-specific heads to capture each class’s semantic essence. Research by You etal. (2022) introduced an innovative incremental transfer learning framework designed to progressively address multi-site medical image segmentation tasks. They proposed strategies at both the model and site levels aimed at incremental training to enhance segmentation, improve generalization, and boost transfer performance, particularly in settings constrained by limited clinical resources.

Ji etal. (2023) introduced an innovative architectural framework for semantic segmentation. Their approach learns a unified deep segmentation model capable of segmenting 143 whole-body organs by training a single unified deep segmentation model. For each segmentation task, a trained encoder remained fixed, whereas a new decoder was gradually added. To handle model complexity, they utilized a progressive pruning strategy involving neural architecture search and teacher-student knowledge distillation. This strategy helps maintain performance while controlling the size of the model.

The practice of dynamically adjusting models extends to classification tasks, as observed in a recent study by Zhang etal. (2023b) on various medical imaging classification tasks. Building upon insights from Ding etal. (2022); Houlsby etal. (2019), they proposed incorporating a lightweight, adaptable module, called an adapter, into a pre-trained and fixed encoder. This addition enables the efficient continual learning of new knowledge by adapting the model for each task to grasp a new set of classes. To address the task identification challenge in continual learning scenarios, task-specific heads were introduced. These heads include an extra class that assimilates all previously learned old classes. Another classification study (Chee etal., 2023) introduced a dynamic architecture featuring expanding representations. Their approach involved incorporating a high-level feature extractor for each task while maintaining a shared low-level feature extractor. This architecture aims to retain common features while learning task-specific characteristics incrementally, asserting the improved preservation of information during each step. In addition, another work by Mousser etal. (2022) proposed an Incremental Deep Tree (IDT) framework tailored for biological image classification. The IDT framework adopted a hierarchical tree-like structure, constructing a new model for each new class by appending branches linked to previously learned tasks. These branches serve to regulate previously acquired knowledge by leveraging replay data and mitigating the issue of catastrophic forgetting. As a result, the framework continuously updates knowledge, enabling the root model to predict new classes while retaining past information and ensuring consistent accuracies for each learned class. Xie etal. (2023) proposes a simple yet effective approach by training batch normalization layer for every task while keeping the rest frozen.

However, while dynamic models offer a logical solution to the problem, identifying the task ID during inference is an open research area for various tasks such as classification and multiclass segmentation. Although dynamic models have proven effective in multi-organ segmentation, this approach may not be optimal, as it requires extensive training and might not be practical. Mitigating forgetting is not the only objective of continual learning.

4 Discussion

Within the medical domain, notable progress has been made in the application of the CL methods. However, the direct use of methods from natural image domains to medical data presents several challenges. Hence, various methodologies have been proposed to address these challenges.

In most studies, there has been a surge in the use of the CL method in medical settings. An ideal and resilient CL model should possess three key characteristics for feasible deployment: maintaining stability, adaptability, and operating within resource constraints. Initially, while training on the first task, the model should be sufficiently flexible to adjust and accommodate new data that may arise in the future. In addition, it is essential for models to retain the knowledge acquired from previous tasks while accommodating new learning tasks. Furthermore, considerations of the size and complexity of the model are important, particularly in deployment scenarios. Therefore, there is a critical need to maintain the model efficiency and scalability to ensure practical feasibility in real-world applications. Moreover, there is a recognized need to develop task-agnostic solutions that are deemed more practical and applicable to real-world scenarios. Consequently, a significant portion of research efforts has been directed towards the development and evaluation of such methodologies.

In this section, we provide a comprehensive overview of the reviewed papers, particularly focusing on CL methods within the medical domain. We identified recent trends, highlighted some of the challenges associated with practical settings, and suggested future directions.

4.1 Trends

In the medical field, the primary focus of tasks centers around segmentation and classification, with relatively less emphasis on detection. The presented figure 7 illustrates a recent growth in both segmentation and classification tasks, with nearly equal attention directed towards both. Segmentation models demand extensive training and computational resources.

Consequently, most of the research has concentrated on regularization approaches to address continual learning challenges. Over half of the reviewed works opted for regularization-based methods to tackle this issue effectively. Furthermore, there has been a recent surge in dynamic model approaches, which, although are less memory efficient and complex to train, offer a simpler solution (vande Ven etal., 2022). Dynamic models, which involve maintaining an ensemble of models for various tasks, offer a straightforward means to handle the problem, contrasting with alternative approaches that may entail intricate learning processes.

Among the different incremental learning scenarios, class-incremental learning stands out as the most challenging and practical setting, where, during inference, there is no specific knowledge available regarding the task of data. Figure 8 demonstrates the significant attention given to the task-agnostic approaches, which is noteworthy. Additionally, most of the literature in this field concentrates on developing methods suitable for task-agnostic settings.

Reproducibility in scientific research is important and is often achieved through the sharing of code and the utilization of publicly available datasets. However, challenges to reproducibility persist, notably when researchers omit details about data splits or utilize private data, thereby hindering accurate replication. Although recent studies commonly utilize publicly available data, many have not shared their code, which significantly hampers reproducibility. Furthermore, most recent research works compare their findings with previous state-of-the-art approaches, providing valuable insights into the field’s progress and establishing benchmarks for future studies. Overall, addressing challenges such as data split information, code sharing, and comparison with previous works is crucial for enhancing the reproducibility and reliability of the research.

Replay and regularization methods in CL prioritize computational efficiency, primarily because of their ability to leverage past experiences or impose constraints on model parameters without necessitating complex dynamic adjustments. Unlike dynamic methods, which often require additional model parameters or extensive memory usage to adapt to evolving tasks, replay methods can efficiently recycle past data samples for training. Similarly, regularization techniques impose constraints on the model weights based on previous tasks, effectively reducing the computational overhead associated with continual adaptation. This focus on computational efficiency ensures smoother integration into real-world applications, where resource constraints are prevalent, enabling CL models to adapt efficiently to new tasks over time.

5 Recommendation

Although significant strides have been made in mitigating catastrophic forgetting within the realm of CL in the medical domain, several critical research avenues remain unexplored, needing attention and investigation in the future. Based on our analysis of the existing literature, the following are some of the open challenges and possible future directions for CL in the medical field.

Specialized CL Datasets in the Medical Domain: There remains a pressing need for specialized datasets tailored to the unique demands of CL scenarios. These datasets would serve as invaluable resources, enabling researchers to advance methodologies within this specialized field. By providing access to data specifically designed for CL tasks, researchers can better understand and address challenges unique to ongoing learning in medical contexts.

Accessible Code Repositories for CL Methodologies: In the ongoing exploration of CL methodologies in the medical field, the availability of easily accessible and replicable code repositories emerges as a critical factor for advancing research endeavors. Accessible code repositories would not only streamline the research process but also facilitate the validation and comparison of various CL approaches. By providing researchers with readily available codebases, the scientific community can foster collaboration and accelerate progress in the development and refinement of CL techniques for medical applications.

Emphasizing Explainability and Interpretability in CL Models: As the development of CL models gains momentum in the medical domain, the emphasis on explainability and interpretability becomes increasingly crucial. In healthcare settings, where transparency and trust are paramount, the ability to understand and interpret model decisions is essential to gaining acceptance from stakeholders. By prioritizing the interpretability of CL models, researchers can enhance their utility and applicability in real-world healthcare contexts, thereby facilitating their integration into clinical practice and decision-making processes.

Practical Deployment Strategies for CL Models in Healthcare: In the endeavor to translate research advancements in CL into tangible benefits for healthcare, the development of practical deployment strategies emerges as a critical consideration. Efforts should be directed towards devising deployment protocols that align with regulatory frameworks and accommodate the timely and appropriate updating of CL models. By addressing deployment challenges such as regulatory compliance and scalability, researchers can bridge the gap between theoretical advancements and their real-world implementation, ultimately facilitating the adoption of CL models in healthcare settings.

6 Conclusion

In practical medical scenarios, there is often a continuous stream of incoming data that requires models to be updated incrementally without encountering issues such as catastrophic forgetting. This paper presents an up-to-date and thorough overview of continual learning applications in the medical domain, encompassing the latest advancements in theory, methodology, and real-world implementation. Our approach emphasizes practicality and deployment considerations. It is promising to note the increasing and widespread interest in continual learning in the medical field, as observed across the broader AI community. This growing interest brings new insights and solutions to address unique medical challenges.

References

• Adlung etal. (2021)Adlung, L., Cohen, Y., Mor, U., Elinav, E., 2021.Machine learning in clinical decision making.Med 2, 642–665.
• Ahsan etal. (2022)Ahsan, M.M., Luna, S.A., Siddique, Z., 2022.Machine-learning-based disease diagnosis: A comprehensive review, in: Healthcare, MDPI. p. 541.
• Akundi and Sivaswamy (2021)Akundi, P., Sivaswamy, J., 2021.Manifold learning to address catastrophic forgetting, in: Proceedings of the Twelfth Indian Conference on Computer Vision, Graphics and Image Processing, pp. 1–5.
• Aljundi etal. (2018)Aljundi, R., Babiloni, F., Elhoseiny, M., Rohrbach, M., Tuytelaars, T., 2018.Memory aware synapses: Learning what (not) to forget, in: Proceedings of the European conference on computer vision (ECCV), pp. 139–154.
• Ayromlou etal. (2022)Ayromlou, S., Abolmaesumi, P., Tsang, T., Li, X., 2022.Class impression for data-free incremental learning, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 320–329.
• Baweja etal. (2018)Baweja, C., Glocker, B., Kamnitsas, K., 2018.Towards continual learning in medical imaging.arXiv preprint arXiv:1811.02496 .
• Bayasi etal. (2023)Bayasi, N., Du, S., Hamarneh, G., Garbi, R., 2023.Continual-gen: Continual group ensembling for domain-agnostic skin lesion classification, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 3–13.
• Bayasi etal. (2021)Bayasi, N., Hamarneh, G., Garbi, R., 2021.Culprit-prune-net: Efficient continual sequential multi-domain learning with application to skin lesion classification, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 165–175.
• Beede etal. (2020)Beede, E., Baylor, E., Hersch, F., Iurchenko, A., Wilcox, L., Ruamviboonsuk, P., Vardoulakis, L.M., 2020.A human-centered evaluation of a deep learning system deployed in clinics for the detection of diabetic retinopathy, in: Proceedings of the 2020 CHI conference on human factors in computing systems, pp. 1–12.
• Bera etal. (2023)Bera, S., Ummadi, V., Sen, D., Mandal, S., Biswas, P.K., 2023.Memory replay for continual medical image segmentation through atypical sample selection, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 513–522.
• Castro etal. (2018)Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K., 2018.End-to-end incremental learning, in: Proceedings of the European conference on computer vision (ECCV), pp. 233–248.
• Chakraborti etal. (2021)Chakraborti, T., Gleeson, F., Rittscher, J., 2021.Contrastive representations for continual learning of fine-grained histology images, in: Machine Learning in Medical Imaging: 12th International Workshop, MLMI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 12, Springer. pp. 1–9.
• Chee etal. (2023)Chee, E., Lee, M.L., Hsu, W., 2023.Leveraging old knowledge to continually learn new classes in medical images.arXiv preprint arXiv:2303.13752 .
• Chen etal. (2023)Chen, B., Thandiackal, K., Pati, P., Goksel, O., 2023.Generative appearance replay for continual unsupervised domain adaptation.arXiv preprint arXiv:2301.01211 .
• Chen etal. (2020)Chen, C., Qin, C., Qiu, H., Tarroni, G., Duan, J., Bai, W., Rueckert, D., 2020.Deep learning for cardiac image segmentation: a review.Frontiers in Cardiovascular Medicine 7, 25.
• Chen etal. (2021)Chen, J., Asma, E., Chan, C., 2021.Targeted gradient descent: A novel method for convolutional neural networks fine-tuning and online-learning, in: Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part III 24, Springer. pp. 25–35.
• Chen and Tang (2022)Chen, S., Tang, F., 2022.Breast cancer detection model training strategy based on continual learning, in: CAIBDA 2022; 2nd International Conference on Artificial Intelligence, Big Data and Algorithms, VDE. pp. 1–5.
• Cohen etal. (2021)Cohen, J.P., Cao, T., Viviano, J.D., Huang, C.W., Fralick, M., Ghassemi, M., Mamdani, M., Greiner, R., Bengio, Y., 2021.Problems in the deployment of machine-learned models in health care.CMAJ 193, E1391–E1394.
• DeFauw etal. (2018)DeFauw, J., Ledsam, J.R., Romera-Paredes, B., Nikolov, S., Tomasev, N., Blackwell, S., Askham, H., Glorot, X., O’Donoghue, B., Visentin, D., etal., 2018.Clinically applicable deep learning for diagnosis and referral in retinal disease.Nature medicine 24, 1342–1350.
• Derakhshani etal. (2022)Derakhshani, M.M., Najdenkoska, I., van Sonsbeek, T., Zhen, X., Mahapatra, D., Worring, M., Snoek, C.G., 2022.Lifelonger: A benchmark for continual disease classification, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 314–324.
• Ding etal. (2022)Ding, N., Qin, Y., Yang, G., Wei, F., Yang, Z., Su, Y., Hu, S., Chen, Y., Chan, C.M., Chen, W., etal., 2022.Delta tuning: A comprehensive study of parameter efficient methods for pre-trained language models.arXiv preprint arXiv:2203.06904 .
• Dosovitskiy etal. (2020)Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., etal., 2020.An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929 .
• Douillard etal. (2021)Douillard, A., Chen, Y., Dapogny, A., Cord, M., 2021.Plop: Learning without forgetting for continual semantic segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 4040–4050.
• Douillard etal. (2022)Douillard, A., Ramé, A., Couairon, G., Cord, M., 2022.Dytox: Transformers for continual learning with dynamic token expansion, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 9285–9295.
• Gao etal. (2023)Gao, J., Xu, L., Wan, M., 2023.Incremental learning for an evolving stream of medical ultrasound images via counterfactual thinking.Computerized Medical Imaging and Graphics 109, 102290.
• van Garderen etal. (2019)van Garderen, K., vander Voort, S., Incekara, F., Smits, M., Klein, S., 2019.Towards continuous learning for glioma segmentation with elastic weight consolidation.arXiv preprint arXiv:1909.11479 .
• González etal. (2022)González, C., Ranem, A., Othman, A., Mukhopadhyay, A., 2022.Task-agnostic continual hippocampus segmentation for smooth population shifts, in: MICCAI Workshop on Domain Adaptation and Representation Transfer, Springer. pp. 108–118.
• Gonzalez etal. (2020)Gonzalez, C., Sakas, G., Mukhopadhyay, A., 2020.What is wrong with continual learning in medical image segmentation?arXiv preprint arXiv:2010.11008 .
• Halder etal. (2020)Halder, A., Dey, D., Sadhu, A.K., 2020.Lung nodule detection from feature engineering to deep learning in thoracic ct images: a comprehensive review.Journal of digital imaging 33, 655–677.
• He etal. (2016)He, K., Zhang, X., Ren, S., Sun, J., 2016.Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778.
• He etal. (2021)He, W., Wang, X., Wang, L., Huang, Y., Yang, Z., Yao, X., Zhao, X., Ju, L., Wu, L., Wu, L., etal., 2021.Incremental learning for exudate and hemorrhage segmentation on fundus images.Information Fusion 73, 157–164.
• Hofmanninger etal. (2020a)Hofmanninger, J., Perkonigg, M., Brink, J.A., Pianykh, O., Herold, C., Langs, G., 2020a.Dynamic memory to alleviate catastrophic forgetting in continuous learning settings, in: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part II 23, Springer. pp. 359–368.
• Hofmanninger etal. (2020b)Hofmanninger, J., Prayer, F., Pan, J., Röhrich, S., Prosch, H., Langs, G., 2020b.Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem.European Radiology Experimental 4, 1–13.
• Houlsby etal. (2019)Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., DeLaroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S., 2019.Parameter-efficient transfer learning for nlp, in: International Conference on Machine Learning, PMLR. pp. 2790–2799.
• Huang etal. (2017)Huang, G., Liu, Z., Van DerMaaten, L., Weinberger, K.Q., 2017.Densely connected convolutional networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4700–4708.
• Huang etal. (2023)Huang, Y., Yang, X., Huang, X., Zhou, X., Chi, H., Dou, H., Hu, X., Wang, J., Deng, X., Ni, D., 2023.Fourier test-time adaptation with multi-level consistency for robust classification.arXiv preprint arXiv:2306.02544 .
• Isensee etal. (2021)Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H., 2021.nnu-net: a self-configuring method for deep learning-based biomedical image segmentation.Nature methods 18, 203–211.
• Ji etal. (2023)Ji, Z., Guo, D., Wang, P., Yan, K., Lu, L., Xu, M., Wang, Q., Ge, J., Gao, M., Ye, X., etal., 2023.Continual segment: Towards a single, unified and non-forgetting continual segmentation model of 143 whole-body organs in ct scans, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 21140–21151.
• Jovicich etal. (2006)Jovicich, J., Czanner, S., Greve, D., Haley, E., van DerKouwe, A., Gollub, R., Kennedy, D., Schmitt, F., Brown, G., MacFall, J., etal., 2006.Reliability in multi-site structural mri studies: effects of gradient non-linearity correction on phantom and human data.Neuroimage 30, 436–443.
• Karani etal. (2018)Karani, N., Chaitanya, K., Baumgartner, C., Konukoglu, E., 2018.A lifelong learning approach to brain mr segmentation across scanners and protocols, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 476–484.
• Kim etal. (2018)Kim, H.E., Kim, S., Lee, J., 2018.Keep and learn: Continual learning by constraining the latent space for knowledge preservation in neural networks, in: Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part I, Springer. pp. 520–528.
• Kim etal. (2024)Kim, J., Lim, M.H., Kim, K., Yoon, H.J., 2024.Continual learning framework for a multicenter study with an application to electrocardiogram.BMC Medical Informatics and Decision Making 24, 67.
• Kirkpatrick etal. (2017)Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., etal., 2017.Overcoming catastrophic forgetting in neural networks.Proceedings of the national academy of sciences 114, 3521–3526.
• Krizhevsky etal. (2012)Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012.Imagenet classification with deep convolutional neural networks.Advances in neural information processing systems 25.
• Kumar etal. (2023)Kumar, R., Kumbharkar, P., Vanam, S., etal., 2023.Medical images classification using deep learning: a survey.Multimedia Tools and Applications doi:10.1007/s11042-023-15576-. published online.
• Lenga etal. (2020)Lenga, M., Schulz, H., Saalbach, A., 2020.Continual learning for domain adaptation in chest x-ray classification, in: Medical Imaging with Deep Learning, PMLR. pp. 413–423.
• Li etal. (2022a)Li, K., Yu, L., Heng, P.A., 2022a.Domain-incremental cardiac image segmentation with style-oriented replay and domain-sensitive feature whitening.IEEE Transactions on Medical Imaging 42, 570–581.
• Li etal. (2022b)Li, M., Cong, Y., Liu, Y., Sun, G., 2022b.Class-incremental gesture recognition learning with out-of-distribution detection, in: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE. pp. 1503–1508.
• Li etal. (2022c)Li, X., Li, H., Ma, L., 2022c.Continual learning of medical image classification based on feature replay, in: 2022 16th IEEE International Conference on Signal Processing (ICSP), IEEE. pp. 426–430.
• Li and Hoiem (2017)Li, Z., Hoiem, D., 2017.Learning without forgetting.IEEE transactions on pattern analysis and machine intelligence 40, 2935–2947.
• Li etal. (2020)Li, Z., Zhong, C., Wang, R., Zheng, W.S., 2020.Continual learning of new diseases with dual distillation and ensemble strategy, in: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part I 23, Springer. pp. 169–178.
• Liao etal. (2022)Liao, W., Xiong, H., Wang, Q., Mo, Y., Li, X., Liu, Y., Chen, Z., Huang, S., Dou, D., 2022.Muscle: Multi-task self-supervised continual learning to pre-train deep models for x-ray images of multiple body parts, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 151–161.
• Liu etal. (2022a)Liu, M., Xiao, L., Jiang, H., He, Q., 2022a.A new generative replay approach for incremental class learning of medical image for semantic segmentation, in: Proceedings of the 2022 International Conference on Intelligent Medicine and Health, pp. 51–56.
• Liu etal. (2022b)Liu, P., Wang, X., Fan, M., Pan, H., Yin, M., Zhu, X., Du, D., Zhao, X., Xiao, L., Ding, L., etal., 2022b.Learning incrementally to segment multiple organs in a ct image, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 714–724.
• Liu etal. (2023)Liu, X., Shih, H.A., Xing, F., Santarnecchi, E., ElFakhri, G., Woo, J., 2023.Incremental learning for heterogeneous structure segmentation in brain tumor mri, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 46–56.
• Morgado etal. (2021)Morgado, A.C., Andrade, C., Teixeira, L.F., Vasconcelos, M.J.M., 2021.Incremental learning for dermatological imaging modality classification.Journal of Imaging 7, 180.
• Moskalenko and Kharchenko (2024)Moskalenko, V., Kharchenko, V., 2024.Resilience-aware mlops for ai-based medical diagnostic system.Frontiers in Public Health 12, 1342937.
• Mousser etal. (2022)Mousser, W., Ouadfel, S., Taleb-Ahmed, A., Kitouni, I., 2022.Idt: An incremental deep tree framework for biological image classification.Artificial Intelligence in Medicine 134, 102392.
• Özgün etal. (2020)Özgün, S., Rickmann, A.M., Roy, A.G., Wachinger, C., 2020.Importance driven continual learning for segmentation across domains, in: Machine Learning in Medical Imaging: 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Proceedings 11, Springer. pp. 423–433.
• Patra etal. (2021)Patra, A., Cai, Y., Chatelain, P., Sharma, H., Drukker, L., Papageorghiou, A.T., Noble, J.A., 2021.Multimodal continual learning with sonographer eye-tracking in fetal ultrasound, in: Simplifying Medical Ultrasound: Second International Workshop, ASMUS 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, September 27, 2021, Proceedings 2, Springer. pp. 14–24.
• Perkonigg etal. (2021)Perkonigg, M., Hofmanninger, J., Herold, C.J., Brink, J.A., Pianykh, O., Prosch, H., Langs, G., 2021.Dynamic memory to alleviate catastrophic forgetting in continual learning with medical imaging.Nature communications 12, 5678.
• Preboske etal. (2006)Preboske, G.M., Gunter, J.L., Ward, C.P., JackJr, C.R., 2006.Common mri acquisition non-idealities significantly impact the output of the boundary shift integral method of measuring brain atrophy on serial mri.Neuroimage 30, 1196–1202.
• Qazi etal. (2023)Qazi, M.A., Alam, M.T., Almakky, I., Diehl, W.G., Bricker, L., Yaqub, M., 2023.Multi-task learning approach for unified biometric estimation from fetal ultrasound anomaly scans, in: International Conference on Medical Imaging and Computer-Aided Diagnosis, Springer. pp. 52–61.
• Qazi etal. (2024)Qazi, M.A., Almakky, I., Hashmi, A.U.R., Sanjeev, S., Yaqub, M., 2024.Dynammo: Dynamic model merging for efficient class incremental learning for medical images.arXiv preprint arXiv:2404.14099 .
• Quarta etal. (2022)Quarta, A., Bruno, P., Calimeri, F., etal., 2022.Continual learning for medical image classification, in: CEUR Workshop Proceedings (CEUR-WS. org).
• Ranem etal. (2022)Ranem, A., González, C., Mukhopadhyay, A., 2022.Continual hippocampus segmentation with transformers, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3711–3720.
• Ravishankar etal. (2019)Ravishankar, H., Venkataramani, R., Anamandra, S., Sudhakar, P., Annangi, P., 2019.Feature transformers: privacy preserving lifelong learners for medical imaging, in: Medical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part IV 22, Springer. pp. 347–355.
• Rebuffi etal. (2017)Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H., 2017.icarl: Incremental classifier and representation learning, in: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp. 2001–2010.
• Roy etal. (2023)Roy, K., Moghadam, P., Harandi, M., 2023.L3dmc: Lifelong learning using distillation via mixed-curvature space, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 123–133.
• Schlegl etal. (2019)Schlegl, T., Seeböck, P., Waldstein, S.M., Langs, G., Schmidt-Erfurth, U., 2019.f-anogan: Fast unsupervised anomaly detection with generative adversarial networks.Medical image analysis 54, 30–44.
• Shevchyk etal. (2022)Shevchyk, A., Hu, R., Thandiackal, K., Heizmann, M., Brunschwiler, T., 2022.Privacy preserving synthetic respiratory sounds for class incremental learning.Smart Health 23, 100232.
• Shu etal. (2022)Shu, K., Li, H., Cheng, J., Guo, Q., Leng, L., Liao, J., Hu, Y., Liu, J., 2022.Replay-oriented gradient projection memory for continual learning in medical scenarios, in: 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE. pp. 1724–1729.
• Smith etal. (2023)Smith, J.S., Karlinsky, L., Gutta, V., Cascante-Bonilla, P., Kim, D., Arbelle, A., Panda, R., Feris, R., Kira, Z., 2023.Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11909–11919.
• Srivastava etal. (2021)Srivastava, S., Yaqub, M., Nandakumar, K., Ge, Z., Mahapatra, D., 2021.Continual domain incremental learning for chest x-ray classification in low-resource clinical settings, in: MICCAI Workshop on Domain Adaptation and Representation Transfer, Springer. pp. 226–238.
• Thandiackal etal. (2023)Thandiackal, K., Piccinelli, L., Pati, P., Goksel, O., 2023.Multi-scale feature alignment for continual learning of unlabeled domains.arXiv:2302.01287.
• Tian etal. (2022)Tian, M., Yang, Q., Gao, Y., 2022.Multi-scale multi-task distillation for incremental 3d medical image segmentation, in: European Conference on Computer Vision, Springer. pp. 369--384.
• vande Ven etal. (2022)vande Ven, G.M., Tuytelaars, T., Tolias, A.S., 2022.Three types of incremental learning.Nature Machine Intelligence 4, 1185--1197.
• Venkataramani etal. (2019)Venkataramani, R., Ravishankar, H., Anamandra, S., 2019.Towards continuous domain adaptation for medical imaging, in: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), IEEE. pp. 443--446.
• Walsh etal. (2018)Walsh, S.L., Calandriello, L., Silva, M., Sverzellati, N., 2018.Deep learning for classifying fibrotic lung disease on high-resolution computed tomography: a case-cohort study.The Lancet Respiratory Medicine 6, 837--845.
• Wang etal. (2023)Wang, L., Zhang, X., Su, H., Zhu, J., 2023.A comprehensive survey of continual learning: Theory, method and application.arXiv preprint arXiv:2302.00487 .
• Wang and Luo (2022)Wang, S., Luo, X., 2022.Incremental learning method for lung nodule detection based on ewc and feature distillation, in: International Conference on Biomedical and Intelligent Systems (IC-BIS 2022), SPIE. pp. 868--874.
• Wang etal. (2021)Wang, Z., Yang, Y., Wen, R., Chen, X., Huang, S.L., Zheng, Y., 2021.Lifelong learning based disease diagnosis on clinical notes, in: Pacific-Asia Conference on Knowledge Discovery and Data Mining, Springer. pp. 213--224.
• Wang etal. (2022a)Wang, Z., Zhang, Z., Ebrahimi, S., Sun, R., Zhang, H., Lee, C.Y., Ren, X., Su, G., Perot, V., Dy, J., etal., 2022a.Dualprompt: Complementary prompting for rehearsal-free continual learning, in: European Conference on Computer Vision, Springer. pp. 631--648.
• Wang etal. (2022b)Wang, Z., Zhang, Z., Lee, C.Y., Zhang, H., Sun, R., Ren, X., Su, G., Perot, V., Dy, J., Pfister, T., 2022b.Learning to prompt for continual learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 139--149.
• Wei etal. (2023)Wei, B.Q., Chen, J.J., Tseng, Y.C., Kuo, P.T.P., 2023.Representative data selection for efficient medical incremental learning, in: 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 1--4.doi:10.1109/EMBC40787.2023.10341107.
• Xie etal. (2023)Xie, X., Xu, J., Hu, P., Zhang, W., Huang, Y., Zheng, W., Wang, R., 2023.Task-incremental medical image classification with task-specific batch normalization, in: Chinese Conference on Pattern Recognition and Computer Vision (PRCV), Springer. pp. 309--320.
• Yang etal. (2021)Yang, J., Shi, R., Ni, B., 2021.Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis, in: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), IEEE. pp. 191--195.
• You etal. (2022)You, C., Xiang, J., Su, K., Zhang, X., Dong, S., Onofrey, J., Staib, L., Duncan, J.S., 2022.Incremental learning meets transfer learning: Application to multi-site prostate mri segmentation, in: International Workshop on Distributed, Collaborative, and Federated Learning, Springer. pp. 3--16.
• Zhang etal. (2021)Zhang, J., Gu, R., Wang, G., Gu, L., 2021.Comprehensive importance-based selective regularization for continual segmentation across multiple sites, in: Medical Image Computing and Computer Assisted Intervention--MICCAI 2021: 24th International Conference, Strasbourg, France, September 27--October 1, 2021, Proceedings, Part I 24, Springer. pp. 389--399.
• Zhang etal. (2023a)Zhang, J., Gu, R., Xue, P., Liu, M., Zheng, H., Zheng, Y., Ma, L., Wang, G., Gu, L., 2023a.S 3 r: Shape and semantics-based selective regularization for explainable continual segmentation across multiple sites.IEEE Transactions on Medical Imaging .
• Zhang etal. (2022)Zhang, J., Xue, P., Gu, R., Gu, Y., Liu, M., Pan, Y., Cui, Z., Huang, J., Ma, L., Shen, D., 2022.Learning towards synchronous network memorizability and generalizability for continual segmentation across multiple sites, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer. pp. 380--390.
• Zhang etal. (2023b)Zhang, W., Huang, Y., Zhang, T., Zou, Q., Zheng, W.S., Wang, R., 2023b.Adapter learning in pretrained feature extractor for continual learning of diseases.arXiv preprint arXiv:2304.09042 .
• Zhang etal. (2023c)Zhang, Y., Li, X., Chen, H., Yuille, A.L., Liu, Y., Zhou, Z., 2023c.Continual learning for abdominal multi-organ and tumor segmentation, in: International conference on medical image computing and computer-assisted intervention, Springer. pp. 35--45.
• Zhao etal. (2023)Zhao, D., Ye, Z., Zheng, W.S., Wang, R., 2023.Class-specific prompts in vision transformer for continual learning of new diseases, in: 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE. pp. 994--999.
• Zhou etal. (2023)Zhou, D.W., Wang, Q.W., Qi, Z.H., Ye, H.J., Zhan, D.C., Liu, Z., 2023.Deep class-incremental learning: A survey.arXiv preprint arXiv:2302.03648 .