Design and Development of Deep Learning based system for detection of breast cancer from Digital Mammograms

Female breast cancer has become the most often diagnosed type of cancer in the world and has become a major health concern across rural and urban areas in India nowadays. Asian women are more likely to have dense breasts than other women in the world. Dense breast tissue can make it more difficult to detect breast cancer and is also linked to an increased risk of breast cancer. Digital Mammography is a very useful technique for screening and is acknowledged as the most reliable method for detecting breast cancer at an early stage, but its accuracy is limited by the radiologists’ clinical experience.

To solve this problem a novel fully automated deep learning-based breast cancer computer-aided diagnosis system (CAD) enabled with a graphical user interface is proposed. Quantitative research will be conducted focusing on Digital Mammogram pre-processing, image segmentation to detect regions of interest, and classification of these images based on the Breast Imaging-Reporting and Data System (BI-RADS) using deep learning models. We propose to design a novel CAD system based on the digital mammograms available at Kasturba Hospital, Manipal Patient Archival System.

In the pre-processing phase, the raw mammography images collected from the above database will be anonymized, the BI-RADS reports and the histopathology reports will be analyzed and information related to various lesions will be extracted from the reports. The final curated data will be further processed through image resizing, and conversion to other image formats for the purpose of readability by the program. The need for various de-noising and enhancement techniques will be studied using various performance evaluation parameters.

In the segmentation phase, the model will be developed to detect the Region of Interest (ROI) after removing the pectoral muscles from the mediolateral oblique (MLO) view and any other artifacts which will be identified after data collection. Various traditional segmentation models and deep learning-based segmentation models will be studied and the best model based on the performance metrics will be identified. In the classification phase, the segmented image will be split into training, validation, and testing set which will be tested on various deep-learning models.

The best model based on the performance will be used as the final model for the prospective testing. In the prospective testing phase we will validate the CAD in patients who consent to participate in the study and will be undergoing mammography for clinical indications