An Investigative Study of Shallow, Deep and Dense Learning Models for Breast Cancer Detection based on Microcalcifications

Abstract

Early cancer diagnosis, detection and treatment continues to be a mammoth task in because of many challenges such as socio and cultural myths, economic conditions, access to healthcare services, healthcare practices, availability of expert oncologists etc. Mammography is a successful screening method for the breast cancer detection. Mammography captures multiple features like masses, microcalcifications etc. Microcalcifications may indicate breast cancer in its early stages and are considered to play a crucial role in early breast cancer diagnosis. In this paper, we have undertaken an investigative study for breast cancer classification by automated learning from mammography images with microcalcifications. Three types of convolutional neural architectures – shallow (ResNet101), deep (VGG101) and dense (DenseNet101) learning models are employed in this investigative study towards contributing to the objective of rapid and early breast cancer diagnosis. To improve the accuracies of the learning models, the features extracted from microcalcifications have been fed to the learning models. We have experimented with varying hyperparameter setup and have recorded the optimal performances of the three models. It has been observed that among the three models, ResNet101 model demonstrated best performance of 94.2% in benign and malicious cancer classification and also demonstrated best performance in terms of time complexity. The dense model DenseNet101 was more sensitive and specific towards the classification of breast cancer using the microcalcifications. VGG101 performed well and has worked with nearly optimal results as that of ResNet 101 with a value of 93.6%.