Scholarlink

<p>Medical image segmentation is a key component of computer-aided diagnosis, yet existing deep learning models often struggle to generalize across diverse datasets and suffer from catastrophic forgetting when learning incrementally. To address these issues, we propose IL-ResUNet++, an incremental learning-based framework that incorporates residual blocks, Parametric Sigmoid Squeeze-and-Excitation (PSE) blocks, Atrous Spatial Pyramid Pooling (ASPP), and Shift Blocks to enhance feature recalibration, multi-scale context capture, and spatial representation. To enhance the datasets, the copy-paste data augmentation method is utilized. Additionally, we incorporate a Style-oriented Replay Module (SRM), which mitigates catastrophic forgetting by replaying past data through generative style modulation, ensuring knowledge retention while adapting to new datasets. The model is evaluated on CVC-ClinicDB and the 2018 Data Science Bowl datasets, achieving Dice Similarity Coefficients of 0.9571 and 0.9437, respectively. These results demonstrate that IL-ResUNet + + provides accurate and computationally efficient segmentation, making it well-suited for practical clinical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

IL-ResUNet++: an incremental learning-based deep resunet framework for medical image segmentation

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Suja Paulose,
  • D. Veera Vanitha

摘要

Medical image segmentation is a key component of computer-aided diagnosis, yet existing deep learning models often struggle to generalize across diverse datasets and suffer from catastrophic forgetting when learning incrementally. To address these issues, we propose IL-ResUNet++, an incremental learning-based framework that incorporates residual blocks, Parametric Sigmoid Squeeze-and-Excitation (PSE) blocks, Atrous Spatial Pyramid Pooling (ASPP), and Shift Blocks to enhance feature recalibration, multi-scale context capture, and spatial representation. To enhance the datasets, the copy-paste data augmentation method is utilized. Additionally, we incorporate a Style-oriented Replay Module (SRM), which mitigates catastrophic forgetting by replaying past data through generative style modulation, ensuring knowledge retention while adapting to new datasets. The model is evaluated on CVC-ClinicDB and the 2018 Data Science Bowl datasets, achieving Dice Similarity Coefficients of 0.9571 and 0.9437, respectively. These results demonstrate that IL-ResUNet + + provides accurate and computationally efficient segmentation, making it well-suited for practical clinical applications.