Saliency-Enhanced Content-Based Image Retrieval for Diagnosis Support in Dermatology Consultation: Reader Study

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Medical content image retrieval aids dermatologists in skin lesion diagnosis.
• Current methods lack interpretability, hindering clinical adoption.

Purpose of the Study:

• Develop an interpretable image retrieval architecture for dermoscopic imaging using saliency maps.
• Enhance feature extraction for improved retrieval performance and clinical trust.

Main Methods:

• Implemented saliency-enhanced content-based image retrieval (SE-CBIR) using convolutional neural networks (CNNs) and saliency maps.
• Conducted a reader study at University Hospital Zurich Dermatology Clinic to assess SE-CBIR's accuracy and impact on diagnostic confidence.

Main Results:

• SE-CBIR improved retrieval accuracy by 7% (77% to 84%) compared to traditional CBIR.
• Reader study showed a 22% increase in classification accuracy and a 24% rise in diagnostic confidence.
• SE-CBIR usage reduced overdiagnosis of nonmelanoma lesions as melanoma by 53%.

Conclusions:

• SE-CBIR offers superior retrieval accuracy and interpretability over traditional CNN-based CBIR.
• This interpretable tool enhances dermatologists' diagnostic accuracy and confidence.
• Expected to increase adoption of AI support tools in routine dermatological consultations.