Machine Learning Augmented Breast Tumors Classification using Magnetic Resonance Imaging Histograms

At present, breast cancer survival rate significantly varies with the stage at which it was first detected. It is crucial to achieve early detection of malignant tumors to reduce their negative effects. Magnetic resonance imaging (MRI) is currently an important imaging modality in the detection of breast tumors. A need exists to develop computer aided methods to provide early diagnosis of malignancy. In this study, I present machine learning models utilizing new image histogram features using the pixels least significant bit. The models were first trained on an MRI breast dataset that included 227 images captured using the short TI inversion recovery (STIR) sequence and diagnosed as either benign or malignant. Three data classification methods were utilized to differentiate between the tumor's classes. The examined classification methods were the Discriminant Analysis, K-Nearest Neighborhood, and the Random Forest. Algorithms' testing was performed on a completely different dataset that included another 186 MRI STIR images showing breast tumors with verified biopsy diagnostics. A significant tumor classification efficiency was found, as judged by the pathological diagnosis. Classification's accuracy was calculated as 94.1% for the DA, 94.6% for the KNN and 80.6% for the RF algorithm. Receiver operating curves also showed significant classification performances. The proposed tumor classification techniques can be used as non-invasive and fast diagnostic tools for breast tumors, with the capability of significantly reducing false errors associated with common MRI imaging-based diagnosis.