Breast cancer detection in thermograms using a hybrid of GA and GWO based deep feature selection method

Breast cancer is one of the most common reasons for the premature death of women worldwide. However, early detection and diagnosis of the same can save many lives. Hence, computer scientists across the world are also striving to develop reliable models to deal with this disease. One of the reliable methods of detecting breast cancer is through thermal images or thermograms. In medical image analysis, the current research trend is the use of deep learning (DL) models. However, many state-of-the-art DL models generate a large number of features and for processing those features we need a significant amount of memory and computation time. To address this issue, we have proposed a lightweight model to detect signs of abnormality in breast thermograms using a combination of transfer learning-based DL model and feature selection approaches. At first, we employed a deep learning model, called SqueezeNet 1.1 (pre-trained on the ImageNet dataset), fine-tuned on the breast cancer thermal images for the feature extraction purpose. Then a hybrid of Genetic Algorithm (GA) and Grey Wolf Optimizer (GWO) is used to reduce the dimension of the obtained feature vector. Before that, a chaotic map is used to create the initial population of GA. The proposed model performs very well in detecting and differentiating malignant and healthy breasts. We have evaluated our model on a publicly available dataset, namely Database For Mastology Research (DMR-IR) and achieved 100% accuracy on the test set using only 3% features extracted by the SqueezeNet 1.1 model.