Deep Learning-based selection of Magnetic Resonance Imaging dataframe in presence of brain cancer

Brain cancer is one of the most frequently occurring type of cancer, accounting for the top-ten cause of mortality all over the world. It is an aggressive malignancy, thus the early detection is of crucial importance for patients’ survival and quality of life. In the current clinical practice, the diagnosis of brain cancer from Magnetic Resonance Imaging (MRI) scans is a very challenging task. Medical experts routinely devote a very high amount of time to select abnormal MRI slices from normal ones by using a manual approach. Thus, an automatic classification procedure may be necessary in this context. This works aims to develop a Deep Learning (DL) method for the classification of MRI slices in presence of brain cancer. For this purpose, the VGG-16 architecture has been exploited as classifier. The images were classified into two different classes: normal and abnormal. This approach was validated on two different scenarios: tumor slices with the substitution of non-tumor slices by totally black images and original MRI containing both normal and abnormal slices. These experiments represent a pivotal point in this study, as they allow to evaluate the model performance in two different contexts. The first trial reported an overall accuracy of 98.19%, sensitivity of 100%, specificity of 96.38%, and F1-score of 98.22%. The second trial, instead, resulted in an overall accuracy of 57.67%, sensitivity of 100%, specificity of 0%, and F1-score of 73.15%. The comparative results demonstrated that guiding the model towards relevant features and discriminating patterns contributes most to cancer identification.