Deep-ReAP: Deep Representations And Partial label learning for Multi-pathology Classification.

Area of Science:

• Medical image analysis
• Artificial intelligence in diagnostics
• Deep learning for pathology detection

Background:

• Automated detection of multiple pathologies in medical images is challenging due to data imbalance and patient-specific variations.
• Existing automated systems struggle with the complexity of co-existing pathologies and diverse presentations.

Purpose of the Study:

• To develop a novel method for creating minimal, optimally trained datasets for deep learning models.
• To enable explainable classification and detection of multiple, co-existing pathologies in medical images.

Main Methods:

• Implemented partial label learning with 1% false labels to identify under-trained pathology categories.
• Fine-tuned deep representations to enhance classification accuracy and explainability.
• Identified an optimal subset of 54% of training images for model training.

Main Results:

• Achieved explainable classification for up to 7 co-existing pathological categories in retinal images.
• The model demonstrated high performance with overall precision/recall/Fβ scores of 57%/87%/80%.
• Successfully handled 36 various combinations of co-existing pathologies.

Conclusions:

• The proposed method effectively reduces dataset size while maintaining high performance for multi-pathology detection.
• This approach facilitates explainable inferencing for multi-label medical image datasets.
• The method addresses key challenges in automated medical diagnostics, particularly data imbalance and variability.