GoogLeNet/DenseNet-201 to classify near-infrared (NIR) spectrum graphs for cancer diagnosis – using pretrained image networks for medical spectroscopy

The study compares sensitivity/specificity of classification by pretrained image networks and traditional Machine Learning (ML) methods. One hundred seven spectra each of benign skin conditions actinic keratosis (ACK) and seborrheic keratosis (SEK), and skin cancer basal cell carcinoma (BCC) were downloaded from a public database. Eighty spectra per group were used for training and twenty-seven spectra per group for testing. In the first classification strategy, spectrum intensity values were used as input for Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Decision Tree (DT), TreeBagger, Ensemble method, Naïve Bayes, Support Vector Machine (SVM), and Artificial Neural Network (ANN). The second strategy involved using spectrum graphs saved as images to train GoogLeNet, Places-365 GoogLeNet, ResNet-50, Inception-V3, DenseNet-201, and NasNetMobile. Strategy 2 yielded better sensitivity/specificity – 0.7/ 0.91 (ACK), 0.7/0.83 (BCC), and 0.63/0.85 (SEK) compared to strategy 1–0.52/0.94 (ACK), 0.7/0.8 (BCC), and 0.5/0.8 (SEK). Grad-CAM mapping suggested that 1100–1200,1350–1450, and 1600–1700 1/cm to be responsible for classification by strategy 2. When these regions plotted as subplots and saved as images were used for training using strategy 2, sensitivity for BCC increases to 0.78. Results suggest using pretrained image networks to classify spectra may yield better results, give a visual understanding of the basis of classification, and provide means to improve classification further.