DEEP AUTOMATIC ALIGNMENT OF MPOX DERMATOLOGICAL HAND PHOTOGRAPHY

Area of Science:

• Medical imaging
• Computer vision
• Dermatology

Background:

• Mpox (mpox) is a viral illness characterized by significant skin manifestations.
• Accurate tracking of mpox lesion development is essential for assessing disease progression and recovery.
• Current methods for lesion monitoring may lack the precision needed for detailed analysis.

Purpose of the Study:

• To develop and evaluate a novel deep learning application for automated monitoring of mpox lesion progression.
• To adapt the VoxelMorph framework for 2D dermatological photographic data to achieve key point alignment in serial images.
• To assess the performance and generalizability of the proposed lesion monitoring method.

Main Methods:

• A deep learning neural network model was trained on a dataset of 1,658 hand photographs.
• The model was evaluated on a test set of 254 images and validated on an additional 500 images, including those with mpox infections.
• Key point alignment and lesion center registration were explored using an adapted VoxelMorph framework for 2D image data.

Main Results:

• The study demonstrated modest but statistically significant improvements in key point and lesion center registration.
• Performance was evaluated across various regularization strengths, indicating potential for precise lesion tracking.
• The method's generalizability was confirmed with a diverse supplementary image set.

Conclusions:

• The developed deep learning approach offers a promising tool for automated tracking of mpox lesion progression.
• While effective, challenges remain due to the complex anatomical structure of hands, particularly in interdigital areas.
• Further refinement is necessary for optimal application in clinical settings, especially for lesions in complex anatomical regions.