Skin Cancer Detection by Laser-Induced Breakdown Spectroscopy Employing Deep Neural Network Based Classification Methods

Abstract

Laser-induced breakdown spectroscopy (LIBS) with deep neural network (DNN) was conducted to classify four types of skin tissue: melanoma, normal, melanocytic nevus and basal cell carcinoma (BCC). Total 1830 LIBS spectra were acquired for 39 lesions excised from 26 patients. LIBS measurements were performed using a 532 nm wavelength laser (τ = 3 ns) and a 6-channel charge-coupled device (CCD) detector. Linear discriminant analysis (LDA) was performed to classify the four samples into six categories, confirming that they could be classified with high accuracy. (Melanoma vs. Normal skin : 100%, Melanoma vs. Melanocytic nevus : 98.5%, Melanoma vs. BCC : 98%, Normal skin vs. melanocytic nevus : 97%, Normal vs. BCC : 89%, Melanocytic nevus vs. BCC : 88%) The elements classifying each sample were identified by Jensen-Shannon divergence (JSD). It was confirmed that melanoma has a higher Mg concentration than normal and melanocytic nevus, which are benign tumors. DNN was then used to classify the four types and to implement a more accurate model, noise from the CCD was added to the raw data and the model was built using 10 times the data through data augmentation. The Micro f1 score for the four classifications was confirmed to be 0.942 and for the three classifications except BCC it was confirmed to be 0.945. The results confirmed that the LIBS classification method can diagnose skin cancer faster and more accurately than the existing method.