EFCNet enhances the efficiency of segmenting clinically significant small medical objects

Area of Science:

• Medical Image Analysis
• Deep Learning
• Biomarker Segmentation

Background:

• Accurate segmentation of small hyperreflective dots in medical imaging is crucial for diagnosing and monitoring diseases like macular edema.
• Existing segmentation models, including Convolutional Neural Networks (CNNs) and Transformers, often fail to capture these minute structures effectively due to information loss.

Purpose of the Study:

• To introduce EFCNet, a novel deep learning model designed for efficient and accurate segmentation of small hyperreflective dots.
• To enhance feature fusion and hierarchical guidance in segmentation models for improved performance on small objects.

Main Methods:

• Developed EFCNet, incorporating a Cross-Stage Axial Attention (CSAA) module for feature fusion and a Multi-Precision Supervision (MPS) module for hierarchical guidance.
• Evaluated EFCNet on two datasets: S-HRD (retinal OCT scans for macular edema) and S-Polyp (colonoscopy images).

Main Results:

• EFCNet significantly outperformed state-of-the-art models on both datasets.
• Achieved average Dice Similarity Coefficient (DSC) gains of 4.88% on S-HRD and 3.49% on S-Polyp.
• Demonstrated superior performance in segmenting smaller objects, where conventional models typically underperform, with Intersection over Union (IoU) improvements of 3.77% and 3.25% respectively.

Conclusions:

• EFCNet offers superior performance in segmenting small, critical biomarkers compared to existing models.
• The novel CSAA and MPS modules contribute to EFCNet's effectiveness, particularly for challenging small object segmentation.
• EFCNet shows significant potential for clinical application in disease diagnosis and monitoring.