DEEP LEARNING EVALUATION OF THERAPEUTIC EFFICACY IN BREAST CANCER FROM RADIOLOGICAL IMAGES

Breast cancer is the leading cause of cancer-related death in women world-wide. Among all the malignant diseases, it accounts for 23% of all cancer deaths. The highest incidence occurs in women over age 50. The heterogeneity of breast carcinoma can arise from neoplastic transformations in myoepithelial or epithelial cells, or even from a stem cell that has the ability to change its nature. Breast cancer can be classified by different criteria regarding histopathological type, grade, stage and receptor status. It is generally diagnosed through screening or in response to the presence of symptoms that prompt further diagnostic examinations. Imaging techniques are the most used for tumor detection and the principal ones are: mammography, ultrasound, positron emission tomography and magnetic resonance imaging. Specifically, tumor identification in breast MRI images relies on the different contrast enhancement between normal tissues and breast lesions. Among all different types of tumors treatments, surgical resection and chemoradiotherapy are often the principal ones. Recently, new therapies that reactivate immune responses against cancer, have discovered the presence of specific killing lymphocytes in the tumor microenvironment, the so called tumor-infiltrating lymphocytes (TIL). The presence of TIL in breast cancer prior to treatment can predict the response to therapy and it is associated with a better prognosis. In recent years, there has been significant progress in Deep Learning (DL) techniques, particularly in the automatic analysis of radiological images for tumor detection and the prediction of therapeutic efficacy. The aim of this study is to assess TIL from breast cancer MRIs, acquired from the Cancer Genome Atlas Breast Invasive Carcinoma data collection. The study relies on three experiments performed using three different classifiers: the VGG-16 and two simpler CNNs, one with a single convolutional layer and the other with four convolutional layers. All these models were fed with MRIs pre-processed with the following customized pipeline: normalization, bias field correction and inter-slice distance modification. Afterward, the hyper-parameter tuning was utilized to find the best parameters to train, validate and test the models. The performances of the three model architectures were compared in terms of accuracy, sensitivity, specificity, area under (AUC) the curve and F1-score. It turned out that the model of intermediate complexity reached better performances with AUC of 0.943, 0.962, 0.990, 0.985 and 0.988 in different splits. The achieved results demonstrate that the classifier used in the third experiment is able to accurately discriminate among low TIL and high TIL tumors from radiological images. This is a very promising starting point for the development of a clinically useful computer aided system to assess TIL levels from radiological images.