Class incremental medical image segmentation (CIMIS) aims to preserve knowledge of previously learned classes while learning new ones without relying on old-class annotations. However, existing methods 1) either adopt one-size-fits-all strategies that treat all spatial regions and feature channels equally, which may hinder the preservation of accurate old knowledge, 2) or focus solely on aligning local prototypes with global ones for old classes while overlooking their local representations in new data, leading to knowledge degradation. To mitigate the above issues, we propose Prototype-Guided Calibration Distillation (PGCD) and Dual-Aligned Prototype Distillation (DAPD) for CIMIS in this paper. Specifically, PGCD exploits prototype-to-feature similarity to calibrate class-specific distillation intensity in different spatial regions, effectively reinforcing reliable old knowledge and suppressing misleading cues from old classes. Complementarily, DAPD aligns the local prototypes of old classes extracted from the current model with both global historical prototypes and local prototypes, further enhancing segmentation performance on old categories. Comprehensive evaluations on two widely used multi-organ segmentation benchmarks demonstrate that our method outperforms current state-of-the-art methods, highlighting its robustness and generalization capabilities.

Inverse treatment planning is pivotal in tumor treatment planning. It enables the multi-objective optimization of radiation dose delivery, ensuring precise tumor targeting while sparing surrounding healthy tissues. This process often requires frequent parameter adjustments to achieve the desired balance between objectives, making it both labor-intensive and time-consuming. Deep reinforcement learning (DRL) provides an automated, model-based planning solution, aimed at reducing reliance on human expertise and enhancing the efficiency of objective parameter optimization. However, most current approaches apply DRL to inverse planning without fully leveraging the knowledge embedded in the continuous state-action space, defined by the coupling between nonstationary planning states and continuous decision variables. This may result in insufficient exploration and exploitation, leading to inefficient optimization. This work introduces an innovative action-guided DRL (AgDRL) approach for automatic inverse planning. Our goal is to enhance exploration and exploitation by leveraging insightful guidance from reward-guided actions. The implementation of AgDRL incorporates both exploitation and exploration in the action-state space. For exploitation, high-reward actions are employed as guidance to achieve the optimal action adjustment. For exploration, low-reward actions are recommended as training resets to explore a broader range of the latent state space. Quantitative and qualitative experiments are conducted in various settings to evaluate the proposed method. The results are assessed using DRL-related metrics (e.g. reward gains) and clinical-related measurements (e.g. dose-volume histograms, DVHs). Experimental results on a real-world rectal cancer dataset empirically demonstrate that the proposed AgDRL-based approach significantly improves optimization efficiency through a high-reward strategy while enhancing exploration diversity via a low-reward strategy, consistently outperforming the MatRad treatment planning optimization platform.

Class incremental medical segmentation (CIMS) aims to sequentially learn new classes while preserving knowledge of previously learned categories in the absence of old-class labels. Current methods suffer from performance degradation under class imbalance and require additional segmentation heads to accommodate new categories. Inspired by recent prototype learning that leverages prototypes to achieve robust recognition of new categories under limited-data regimes, we introduce a Prototype-dRIven class increMEntal (PRIME) method. PRIME replaces the incremental segmentation heads with prototypes to mitigate class imbalance, allowing new class learning with the simple addition of new prototypes. Based on prototype learning, PRIME further involves three tailored techniques. First, prototype structure alignment imposes structural constraints on inter-prototype relations to maintain consistent relative distances in the feature space, improving the model's ability to distinguish distinct classes. Second, pixel-wise contrastive loss term groups embeddings of similar samples while separating those of different classes, enhancing segmentation accuracy across all categories. Finally, the consensus-based prototype update mechanism refines the old prototypes during the learning of new classes, preventing performance degradation on the old classes. Extensive experiments on two public multi-organ segmentation datasets demonstrate that our approach significantly outperforms state-of-the-art methods, validating the effectiveness of the proposed PRIME.

How to balance the learning ’sensitivity-stability’ upon new task training and memory preserving is critical in CL to resolve catastrophic forgetting. Improving model generalization ability within each learning phase is one solution to help CL learning overcome the gap in the joint knowledge space. Zeroth-order loss landscape sharpness-aware minimization is a strong training regime improving model generalization in transfer learning compared with optimizer like SGD. It has also been introduced into CL to improve memory representation or learning efficiency. However, zeroth-order sharpness alone could favors sharper over flatter minima in certain scenarios, leading to a rather sensitive minima rather than a global optima. To further enhance learning stability, we propose a Continual Flatness (C-Flat) method featuring a flatter loss landscape tailored for CL. C-Flat could be easily called with only one line of code and is plug-and-play to any CL methods. A general framework of C-Flat applied to all CL categories and a thorough comparison with loss minima optimizer and flat minima based CL approaches is presented in this paper, showing that our method can boost CL performance in almost all cases. Code is available at https://github.com/WanNaa/C-Flat.

Precisely segmenting the organs at risk (OARs) in computed tomography (CT) plays an important role in radiotherapy’s treatment planning, aiding in the protection of critical tissues during irradiation. Renowned deep convolutional neural networks (DCNNs) and prevailing transformer-based architectures are widely utilized to accomplish the segmentation task, showcasing advantages in capturing local and contextual characteristics. Graph convolutional networks (GCNs) are another specialized model designed for processing the nongrid dataset, e.g., citation relationship. The DCNNs and GCNs are considered as two distinct models applicable to the grid and nongrid datasets, respectively. Motivated by the recently developed dynamic-channel GCN (DCGCN) that attempts to leverage the graph structure to enhance the feature extracted by the DCNNs, this paper proposes a novel architecture termed adaptive sparse GCN (ASGCN) to mitigate the inherent limitations in DCGCN from the aspect of node’s representation and adjacency matrix’s construction. For the node’s representation, the global average pooling used in the DCGCN is replaced by the learning mechanism to accommodate the segmentation task. For the adjacency matrix, an adaptive regularization strategy is leveraged to penalize the coefficient in the adjacency matrix, resulting in a sparse one that can better exploit the relationships between nodes. Rigorous experiments on multiple OARs’ segmentation tasks of the head and neck demonstrate that the proposed ASGCN can effectively improve the segmentation accuracy. Comparison between the proposed method and other prevalent architectures further confirms the superiority of the ASGCN.

Deep neural networks (DNNs) have emerged as a prominent model in medical image segmentation, achieving remarkable advancements in clinical practice. Despite the promising results reported in the literature, the effectiveness of DNNs necessitates substantial quantities of high-quality annotated training data. During experiments, we observe a significant decline in the performance of DNNs on the test set when there exists disruption in the labels of the training dataset, revealing inherent limitations in the robustness of DNNs. In this paper, we find that the neural memory ordinary differential equation (nmODE), a recently proposed model based on ordinary differential equations (ODEs), not only addresses the robustness limitation but also enhances performance when trained by the clean training dataset. However, it is acknowledged that the ODE-based model tends to be less computationally efficient compared to the conventional discrete models due to the multiple function evaluations required by the ODE solver. Recognizing the efficiency limitation of the ODE-based model, we propose a novel approach called the nmODE-based knowledge distillation (nmODE-KD). The proposed method aims to transfer knowledge from the continuous nmODE to a discrete layer, simultaneously enhancing the model's robustness and efficiency. The core concept of nmODE-KD revolves around enforcing the discrete layer to mimic the continuous nmODE by minimizing the KL divergence between them. Experimental results on 18 organs-at-risk segmentation tasks demonstrate that nmODE-KD exhibits improved robustness compared to ODE-based models while also mitigating the efficiency limitation.