Abstract
INTRODUCTION: This study presents a novel continuous learning framework tailored for brain tumour segmentation, addressing a critical step in both diagnosis and treatment planning. This framework addresses common challenges in brain tumour segmentation, such as computational complexity, limited generalisability, and the extensive need for manual annotation. METHODS: Our approach uniquely combines multi-scale spatial distillation with pseudo-labelling strategies, exploiting the coordinated capabilities of the ResNet18 and DeepLabV3+ network architectures. This integration enhances feature extraction and efficiently manages model size, promoting accurate and fast segmentation. To mitigate the problem of catastrophic forgetting during model training, our methodology incorporates a multi-scale spatial distillation scheme. This scheme is essential for maintaining model diversity and preserving knowledge from previous training phases. In addition, a confidence-based pseudo-labelling technique is employed, allowing the model to self-improve based on its predictions and ensuring a balanced treatment of data categories. RESULTS: The effectiveness of our framework has been evaluated on three publicly available datasets (BraTS2019, BraTS2020, BraTS2021) and one proprietary dataset (BraTS_FAHZU) using performance metrics such as Dice coefficient, sensitivity, specificity and Hausdorff95 distance. The results consistently show competitive performance against other state-of-the-art segmentation techniques, demonstrating improved accuracy and efficiency. DISCUSSION: This advance has significant implications for the field of medical image segmentation. Our code is freely available at https://github.com/smallboy-code/A-brain-tumor-segmentation-frameworkusing-continual-learning.