AI-driven discrimination of benign from malignant pigmented skin lesions based on multispectral autofluorescence lifetime dermoscopy imaging

Author(s)

Vasanthakumari, Priyanka ; Romano, Renan A. ; Rosa, Ramon G. T. ; Salvio, Ana G. ; Yakovlev, Vladislav ; Kurachi, Cristina ; Jo, Javier A. ; Choi, B ; Zeng, H

• Document type: Journal article
• Source: PHOTONICS IN DERMATOLOGY AND PLASTIC SURGERY 2022; v. 11934, p. 13-pg., 2022-01-01.

Abstract

Melanoma is the most aggressive type of skin cancer with an estimated 106,110 new cases in the US in 2021. The 5-year survival rate of patients with early-stage melanoma is similar to 99%; however, similar to 13% of melanoma patients are diagnosed with lesions already at intermediate or advance stages, associated with a 5-year survival rate of similar to 66% and similar to 27% respectively. The current diagnosis technique involving visual inspection and biopsy often fail to visually distinguish clinically similar lesions; in particular, melanoma can be mistaken for benign lesion pigmented seborrheic keratosis (pSK). In this work, a deep learning model using Long Short-Term Memory (LSTM) networks is trained on the multispectral autofluorescence lifetime dermoscopy images collected from 41 benign lesions including solar lentigo and pSK, and 19 malignant lesions including melanoma, superficial basal cell carcinoma (BCC) and nodular BCC. The model is trained on the image pixels containing concatenated fluorescent decay signals from three emission channels. The posterior probabilities predicted for each pixel location, is used to construct probability maps of the images. Receiver Operator Characteristics (ROC) constructed on the threshold of the median value of the posterior probability map determines the effectiveness in distinguishing benign and malignant lesions. The entire dataset is split into training, validation, and test sets. The hyperparameters are tuned using the validation set while the model performance is estimated using the test set. The mean and standard deviation of the Areas Under the Curve (AUC) of the ROCs generated with 10 random test sets is 0.82 +/- 0.04. (AU)

• FAPESP's process: 14/50857-8 - National Institute in Basic Optics and Applied to Life Sciences
• Grantee: Vanderlei Salvador Bagnato
• Support Opportunities: Research Projects - Thematic Grants
• FAPESP's process: 13/07276-1 - CEPOF - Optics and Photonic Research Center
• Grantee: Vanderlei Salvador Bagnato
• Support Opportunities: Research Grants - Research, Innovation and Dissemination Centers - RIDC