Automated melanoma detection: multispectral imaging and neural network approach for classification

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Cutaneous pigmented lesions require accurate diagnosis to differentiate between melanoma and non-melanoma types.
• Early melanoma detection significantly improves patient prognosis.
• Traditional diagnostic methods have limitations in accurately classifying pigmented skin lesions.

Purpose of the Study:

• To evaluate the diagnostic performance of artificial neural networks (ANNs) using multispectral images of skin lesions.
• To compare the efficacy of ANNs against multivariate discriminant analysis (MDA) for lesion classification.
• To assess the diagnostic accuracy and stability of both AI and traditional methods.

Main Methods:

• A cohort of 573 cutaneous pigmented lesions from 534 patients was analyzed using a telespectrophotometric system (TS) capturing 17 images (400-1040 nm).
• Five lesion descriptors based on ABCD criteria were extracted per wavelength, reduced via factor analysis into ten variables.
• Data were split into training (400 cases) and verification (173 cases) sets for ANN and MDA model development.

Main Results:

• ANNs achieved 80% sensitivity and 72% specificity on the training set, and 78% sensitivity and 76% specificity on the validation set.
• Multivariate discriminant analysis (MDA) yielded 80% sensitivity and 60% specificity (train), and 76% sensitivity and 57% specificity (validation).
• Area Under the Curve (AUC) for ANNs was 0.852 (train) and 0.847 (verify), outperforming MDA's AUC of 0.810 (train) and 0.764 (verify).

Conclusions:

• Artificial neural networks demonstrate superior diagnostic performance and model stability for classifying cutaneous pigmented lesions compared to MDA.
• ANNs applied to multispectral imaging show promise as an advanced tool for melanoma diagnosis.
• The study highlights the potential of AI in improving the accuracy of dermatological assessments.