Deep Collocative Learning for Immunofixation Electrophoresis Image Analysis.

Area of Science:

• Medical Diagnostics
• Artificial Intelligence
• Biomedical Imaging

Background:

• Immunofixation Electrophoresis (IFE) is crucial for Multiple Myeloma diagnosis.
• Deep learning has not been extensively applied to IFE pattern recognition.
• Existing deep learning models struggle with IFE's relational band patterns and lack explainability.

Purpose of the Study:

• To develop a deep learning method for accurate IFE pattern analysis.
• To address the challenges of relational pattern recognition in IFE.
• To enhance the interpretability and validation of deep learning predictions for IFE.

Main Methods:

• Introduced collocative learning to convert IFE's binary relations into unary relations for deep networks.
• Developed a location-label-free method using Grad-CAM for accurate band localization.
• Proposed Coached Attention Gates to align model inference with human logic for evidence backtracking.

Main Results:

• Achieved a 741.30% improvement in Intersection over Union (IoU) compared to the ResNet18 baseline.
• Demonstrated superior performance over established deep networks like DenseNet, CBAM, and Inception-v3.
• Enabled accurate localization and provided verifiable evidence for IFE pattern predictions.

Conclusions:

• The proposed collocative learning framework effectively addresses limitations of deep learning in IFE analysis.
• The method enhances diagnostic accuracy and provides crucial explainability for clinical validation.
• This approach holds significant potential for improving Multiple Myeloma diagnosis through AI.