MAP-SCTNet: multi-scale pyramid and frequency-enhanced network for colorectal cancer histopathological image
Haitao Chen1, Zhengxu Duan2, Lan Li3
1Chengdu Qingbaijiang District People's Hospital, Chengdu, China.
Frontiers in Medicine
|March 30, 2026
Summary
MAP-SCTNet improves medical image segmentation for colorectal cancer detection by integrating multi-scale features and texture analysis. This efficient network enhances diagnostic accuracy with fewer parameters and computations, aiding real-time clinical applications.
Area of Science:
- Medical image analysis
- Computational pathology
- Deep learning for medical imaging
Background:
- Medical image segmentation is crucial for diagnosing diseases like colorectal cancer but faces challenges like lesion heterogeneity, low contrast, and limited data.
- Existing deep learning models often fail to balance multi-scale context and boundary details, especially in resource-limited clinical settings.
Purpose of the Study:
- To develop an efficient segmentation network, MAP-SCTNet, that addresses limitations in medical image segmentation for colorectal cancer histopathology.
- To enhance segmentation accuracy by improving multi-scale feature extraction, boundary preservation, and texture representation.
Main Methods:
- Proposed MAP-SCTNet, incorporating a Multi-Scale Atrous Spatial Pyramid Pooling (MS-ASPP) module with dual-attention for multi-scale and boundary information.
- Introduced an Adaptive Frequency Enhancement and Texture-Aware Module (AFE-TAM) for robust texture analysis across pathological subtypes.
- Implemented a Progressive Dual-Teacher Knowledge Distillation (PDTKD) framework to improve generalization from limited annotated data.
Main Results:
- MAP-SCTNet achieved state-of-the-art performance on the EBHI-SEG colorectal cancer dataset, with a mean Dice coefficient of 83.76% and mean IoU of 74.5%.
- The model significantly outperformed recent methods like TransUNet, Swin-UNet, and SegNeXt.
- MAP-SCTNet demonstrated high efficiency with 21.8M parameters and 30.7G FLOPs, a substantial reduction compared to TransUNet.
Conclusions:
- MAP-SCTNet offers a highly efficient and accurate solution for colorectal cancer histopathological image segmentation.
- The proposed architecture effectively captures multi-scale context, preserves fine boundary details, and robustly represents texture.
- Its computational efficiency makes it suitable for real-time computer-aided diagnosis in clinical workflows.
