Abstract:
In screening mammography for breast cancer, radiologists often identify the region of interest (ROI) in a mammogram, and analyze it whether it's benign or malignant, and whether it's a mass or calcifi cation abnormality, for the course of treatment is dependent on the result of the exam. However, there are cases where the ROI is extremely difficult to classify, and radiologists sometimes have conflicting readings. Consequenty, Mammo, an application that utilized Convolutional Neural Networks (CNN), was created in this study providing radiologists and medical field students/trainees a tool for classifying their mammograms as well as an exam for the students/trainees. The CNN architecture used was ResNet-50, a CNN where layers can be stacked without typical vanishing or exploding gradients problem that plague CNNs when more layers are stacked. Using CBIS-DDSM (Curated Breast Imaging Subset of the Digital Database for Screening Mammography), initially, the network achieved an accuracy of 33%. This is caused by lack of data augmenting technique performed on the dataset as well as imabalance of instances per category. After balancing the instances, CNN ResNet-50 achieved an overall accuracy rate of 43%.
