Diagnostic accuracy of content-based dermatoscopic image retrieval with deep classification features.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Neural networks achieve high accuracy in medical image classification but often lack interpretability.
• Automated diagnostic tools are increasingly important in clinical settings.

Purpose of the Study:

• To compare the diagnostic accuracy of content-based image retrieval (CBIR) with neural network predictions for dermatoscopic images.
• To evaluate CBIR's potential to improve diagnostic accuracy for clinicians.

Main Methods:

• Trained neural networks on three dermatoscopic image datasets (888, 2750, 16,691 images).
• Used CBIR to retrieve visually similar images and predict diagnoses based on common diagnoses or top-1 network prediction.
• Evaluated performance using area under the receiver operating characteristic curve (AUC), multiclass-accuracy, and mean average precision (mAP).

Main Results:

• CBIR demonstrated AUC values comparable to neural network softmax predictions across all datasets (e.g., 0.842 vs. 0.830).
• Multiclass-accuracy of CBIR was also comparable to softmax predictions.
• CBIR outperformed softmax predictions when networks were trained on fewer classes than present in the dataset (mAP 0.368 vs. 0.184).

Conclusions:

• Presenting visually similar images via CBIR achieves accuracy comparable to neural network probability-based diagnoses.
• CBIR may offer advantages over standard softmax classifiers for improving clinician diagnostic accuracy in routine practice.