Early Detection of Breast Cancer Using Thermal Images: A Study with Light Weight Deep Learning Models

The occurrence rate of cancer is gradually expanding worldwide, and early detection is preferred. Breast Cancer (BC) is a medical emergency, and proper detection is needed to reduce its harshness. The clinical-level screening of BC with Thermal Imaging (TI) is widely adopted due to its accurateness. This work presents the examination of the BC using the TIP and the Pre-trained Light Weight Deep Learning (PLWDL) scheme. The implemented procedure involves (i) Image assortment and modification, (ii) Feature removal and Firefly Algorithm (FA)-based feature optimization, (iii) Binary classification, and (iv) Verification of the clinical significance based on achieved results. Due to its simplicity, the gray-scale version of the thermal images is considered for evaluation using the PLWDL schemes, such as SqueezeNet, MobileNetV1, and MobileNetV2. The detection process is executed using binary classification using SoftMax (SM), Naïve Bayes (NB), and Random Forest (RF), and the experimental outcome achieved is that the SqueezeNet with RF classifier delivers a detection accuracy >90%.