Incremental Learning for Dermatological Imaging Modality Classification.

Area of Science:

• Artificial Intelligence in Medicine
• Teledermatology Image Analysis
• Machine Learning for Healthcare

Background:

• Teledermatology adoption necessitates improved medical record organization.
• Dermatological image modality classification is crucial for filtering records.
• Existing medical imaging modality classification lacks dermatology-specific focus.

Purpose of the Study:

• To classify dermatological images by modality using deep learning models.
• To evaluate incremental learning strategies for adapting models to changing acquisition conditions.
• To assess models based on accuracy, catastrophic forgetting, and computational cost.

Main Methods:

• Utilized VGG-16 and MobileNetV2 models for image classification.
• Applied four incremental learning strategies: naive, EWC, AGEM, and experience replay.
• Evaluated performance considering accuracy, forgetting, and resource usage.

Main Results:

• MobileNetV2 with experience replay achieved 86.04% accuracy.
• This strategy demonstrated minimal forgetting (0.0344), outperforming others by 6.98%.
• The approach incurred additional training time (56s/epoch) and RAM (4554MB).

Conclusions:

• The proposed approach effectively classifies dermatological image modalities.
• Incremental learning, particularly experience replay, enables model adaptation without significant knowledge loss.
• This method shows promise for organizing teledermatology records.