Skin lesion image retrieval using transfer learning-based approach for query-driven distance recommendation.

Area of Science:

• Medical Informatics
• Computer Vision
• Dermatology

Background:

• Content-Based Dermatological Lesion Retrieval (CBDLR) systems aid in diagnosing skin lesions by finding similar images with confirmed diagnoses.
• Accurate and timely diagnosis is crucial for patient survival and reducing healthcare costs.
• Existing CBDLR systems often lack dynamic adaptability in similarity metric selection.

Purpose of the Study:

• To propose an advanced CBDLR system integrating a similarity measure recommender for dynamic distance metric selection.
• To leverage deep-learned features for enhanced skin lesion classification.
• To develop an automatic ground truth generation method using transfer learning for optimal distance metric recommendation.

Main Methods:

• Utilized deep-learned features validated for their performance in classifying skin lesions into seven distinct classes.
• Implemented a similarity measure recommender for dynamic selection of appropriate distance metrics.
• Employed transfer learning for automatic ground truth generation to recommend optimal distance metrics for new query images.

Main Results:

• The proposed CBDLR system demonstrated superior performance compared to systems using standard distances.
• Achieved at least a 9% improvement in mAP@K (mean Average Precision at K) on the ISIC2018 and ISIC2019 datasets.
• The system provides a valuable aided-decision support tool for dermatologists.

Conclusions:

• The developed CBDLR system offers significant performance gains through dynamic similarity metric selection and deep-learned features.
• The system effectively supports dermatologists in early skin lesion diagnosis, enhancing patient outcomes.
• The approach shows promise for improving the accuracy and efficiency of computer-aided dermatological diagnosis.