Analyzing the Classification Performance of DenseNet121 on Pre-processsed MIAS Dataset

Breast cancer is the most diagnosed type of cancer as per the data that is collected by the World Health Organization (WHO) within the past few years. Over 600,000 deaths were recorded in 2021 due to breast cancer. Breast cancer screening is done using two-dimensional (2D) and three-dimensional (3D) mammography, but MRIs and Ultrasounds are also used in certain conditions. The diagnosis from the screenings is not always accurate as a practitioner must physically look at the digital images to find any signs of cancer. Approximately, each diagnosis has a variable chance of a false-positive or a false-negative. Many CAD (computer-aided detection) systems have been developed for the assistance of a practitioner with the diagnosis. However, in the past years, Deep Neural Networks (DNN) have seen a spike and the models are being used to aid breast cancer screening. Data shows a possibility of reaching Area under the curve (AUC) values as high as 0.99 under ideal conditions when the training dataset is cleaned of noise and properly pre-processed and in some studies, the accuracy and sensitivity are even compared to that of a practitioner’s, with the DNN model outperforming in numbers across the board. After performing a literature review on similar work, we have trained a model of our own on a publicly available dataset (MIAS) reaching promising results of an AUC of 0.87 and an Accuracy of 0.88 with the initial model built on a DenseNet121 architecture.