Characterizing Malignant Melanoma Clinically Resembling Seborrheic Keratosis Using Deep Knowledge Transfer

Area of Science:

• Dermatology and Computer Science
• Medical Image Analysis
• Artificial Intelligence in Healthcare

Background:

• Seborrheic keratosis-like malignant melanomas (SK-like MMs) pose diagnostic challenges, risking delayed treatment.
• Distinguishing SK-like MMs from benign seborrheic keratosis (SK) is crucial to avoid unnecessary excisions or missed diagnoses.

Purpose of the Study:

• To develop and validate a computer-based diagnostic approach for characterizing SK-like MMs using dermoscopy images.
• To improve the accuracy and efficiency of diagnosing atypical melanomas that mimic benign skin lesions.

Main Methods:

• Utilized dermoscopy images from the International Skin Imaging Collaboration archive.
• Employed a convolutional neural network (VGG16) for image embeddings and trained a support vector machine (SVM) classifier.
• Optimized SVM hyperparameters using Bayesian optimization and tested on an independent dataset.

Main Results:

• The SVM classifier demonstrated a sensitivity of 78.6% and a specificity of 84.5% in identifying atypical MMs.
• Key differentiating features included the presence/absence of pigmented networks and milia-like cysts.

Conclusions:

• Computer-based analysis of dermoscopy images shows promise for characterizing challenging SK-like MMs.
• Deep learning and machine learning models can significantly aid in complex dermatological diagnostic problems.