MAP-SCTNet: multi-scale pyramid and frequency-enhanced network for colorectal cancer histopathological image segmentation

MAP-SCTNet:用于结直肠癌组织病理图像分割的多尺度金字塔和频率增强网络

阅读:2

Abstract

Medical image segmentation remains challenging due to the morphological heterogeneity of lesions, inherently low contrast with ambiguous boundaries, and limited availability of annotated training data. Existing methods often struggle to simultaneously capture multi-scale contextual information and preserve fine-grained boundary details, particularly in resource-constrained clinical settings. To address these challenges, we propose MAP-SCTNet, an efficient segmentation network that enhances the SCTNet architecture through three synergistic innovations. First, we design a Multi-Scale Atrous Spatial Pyramid Pooling (MS-ASPP) module that employs five parallel dilated convolution branches coupled with a dual-attention mechanism and boundary-aware refinement, enabling comprehensive multi-scale feature extraction while maintaining precise edge localization. Second, we introduce an Adaptive Frequency Enhancement and Texture-Aware Module (AFE-TAM) that operates in the frequency domain to adaptively fuse low, mid, and high-frequency components, complemented by Gabor filtering and Local Binary Patterns for robust texture representation across pathological subtypes. Third, we develop a Progressive Dual-Teacher Knowledge Distillation (PDTKD) framework that leverages both Transformer-based and CNN-based teacher networks through a three-stage training strategy, effectively bridging the gap between training and inference while improving generalization from limited data. Extensive experiments on the EBHI-SEG colorectal cancer histopathological dataset, which comprises biopsy specimens collected via colonoscopy, demonstrate that MAP-SCTNet achieves state-of-the-art performance with a mean Dice coefficient of 83.76% and mean IoU of 74.5%, surpassing recent methods including TransUNet, Swin-UNet, and SegNeXt. Notably, our model requires only 21.8M parameters and 30.7G FLOPs, representing a 79.3% reduction in parameters and 71.1% reduction in computation compared to TransUNet, making it well-suited for real-time computer-aided diagnosis in clinical colorectal cancer diagnosis workflows.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。