Pyramid Pixel Context Adaption Network for Medical Image Classification With Supervised Contrastive Learning.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Spatial attention (SA) mechanisms in deep neural networks (DNNs) excel at long-range dependency modeling for computer vision.
• However, SA mechanisms often underperform in medical image analysis, struggling to highlight subtle lesion regions due to limitations in long-range dependency modeling.

Purpose of the Study:

• To address the limitations of existing attention mechanisms in medical image analysis.
• To propose a novel, lightweight architectural unit for improved medical image classification.

Main Methods:

• Introduced the Pyramid Pixel Context Adaption (PPCA) module, which dynamically recalibrates pixel positions using multiscale pixel context information.
• Developed the PPCA network (PPCANet) by integrating the PPCA module into a DNN with minimal overhead.
• Incorporated supervised contrastive learning with contrastive loss (CL) to enhance feature representation using label information.

Main Results:

• PPCANet demonstrated superior performance compared to state-of-the-art (SOTA) attention-based networks and recent DNNs across six medical image datasets.
• The PPCA module effectively aggregates multiscale pixel context, normalizes inconsistencies, and estimates per-pixel attention weights.
• Supervised contrastive learning further boosted feature representation quality.

Conclusions:

• The proposed PPCA module and PPCANet offer a practical and effective solution for medical image classification.
• PPCANet overcomes the limitations of traditional spatial attention in identifying subtle lesions.
• The study provides visual analysis and ablation studies validating PPCANet's decision-making process and effectiveness.