Acral melanoma detection using a convolutional neural network for dermoscopy images

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Acral melanoma is the most common melanoma type in Asian populations.
• Late diagnosis of acral melanoma often leads to a poor prognosis.

Purpose of the Study:

• To evaluate the usefulness of a convolutional neural network (CNN) for early diagnosis of acral melanoma.
• To compare CNN performance against dermatologists and non-experts using dermoscopy images.

Main Methods:

• A dataset of 724 dermoscopy images (350 acral melanoma, 374 benign nevi) was analyzed.
• 2-fold cross-validation was employed, splitting data into training and testing sets.
• Diagnostic accuracy was calculated and compared to expert and non-expert evaluations.

Main Results:

• The CNN achieved diagnostic accuracy of 83.51% and 80.23%.
• CNN performance exceeded non-expert evaluations (67.84%, 62.71%) and neared expert levels (81.08%, 81.64%).
• Area-under-the-curve and Youden's index values for the CNN were comparable to expert performance.

Conclusions:

• Convolutional neural networks can be a valuable tool for detecting acral melanoma from dermoscopy images.
• Further data analysis is recommended to enhance CNN accuracy.
• Early detection of acral melanoma can be improved with AI-assisted diagnostic tools.