An optimized multi-task contrastive learning framework for HIFU lesion detection and segmentation.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Accurate detection and segmentation of High-Intensity Focused Ultrasound (HIFU) lesions are crucial for automated disease diagnosis.
• Traditional methods face challenges due to scarce labeled data and suboptimal performance with data variations.
• Existing approaches often struggle with the limited availability of annotated medical datasets.

Purpose of the Study:

• To introduce an innovative Optimized Multi-Task Contrastive Learning Framework (OMCLF) for enhanced HIFU lesion detection and segmentation.
• To leverage self-supervised learning (SSL) and genetic algorithms (GA) to overcome limitations of traditional methods.
• To integrate classification and segmentation into a unified model for improved feature extraction and optimization.

Main Methods:

• Developed the Optimized Multi-Task Contrastive Learning Framework (OMCLF) integrating classification and segmentation with a shared backbone.
• Employed self-supervised learning (SSL) and genetic algorithms (GA) for optimizing feature representations, hyperparameters, and data augmentation.
• Systematically optimized augmentation strategies tailored for medical imaging to preserve critical lesion details.

Main Results:

• OMCLF achieved 93.3% accuracy in lesion detection and a 92.5% Dice score in segmentation.
• The framework outperformed single-task methods and state-of-the-art approaches like SimCLR and MoCo.
• Demonstrated significant reduction in dependency on labeled data for HIFU lesion analysis.

Conclusions:

• OMCLF represents a substantial advancement in automated diagnosis for HIFU-induced lesions.
• The proposed approach offers superior accuracy in identifying and delineating lesions in medical imaging.
• OMCLF signifies a significant step in the evolutionary optimization of SSL for diverse medical imaging applications.