Background selection schema on deep learning-based classification of dermatological disease.

Area of Science:

• Dermatology
• Artificial Intelligence
• Computer Vision

Background:

• Skin diseases are common, and deep learning aids diagnosis but struggles with imbalanced datasets.
• Imbalanced background information in dermatological datasets hinders deep learning model performance.
• Improving skin disease classification requires addressing data imbalances and background variations.

Purpose of the Study:

• To investigate the impact of color-based background manipulation on deep learning models for skin disease classification.
• To enhance the model's ability to learn foreground lesion attributes by controlling background information.
• To identify optimal background colors for improving classification accuracy in dermatological AI.

Main Methods:

• Clinical photographs were annotated, and backgrounds were masked with distinct colors to create varied datasets.
• Data augmentation and random over/undersampling techniques were used to balance training and test sets.
• Deep learning networks (residual networks) were trained independently on subsets with different background colors.

Main Results:

• Color-based background information significantly influences skin disease classification performance.
• Classifiers trained on the green background subset achieved state-of-the-art results.
• The green subset demonstrated superior performance for classifying black and red skin lesions.

Conclusions:

• Strategic background manipulation is crucial for enhancing deep learning models in dermatology.
• Green background masking offers a promising approach to improve skin lesion classification accuracy.
• Further research into background effects can optimize AI diagnostic tools for skin conditions.