Challenges pertaining to Software Engineering in Health Care Systems (A Case Study on Deep Learning based Mammogram Classification)

Abstract

For us being healthy is the most important aspect for longevity. So, better health-care systems need to be developed for diagnosis of diseases. Software Engineering plays an important role in developing health care systems. But while connecting software engineering directly in health care systems, it creates many challenges. So, a case study has been taken i.e. breast cancer. It is the second most dangerous cancer which affect women with increasing death rate. So, for reducing death rate, early recognition and treatment are very important. Mammography is an efficient way of finding breast cancer. But reading of same mammogram by different radiologists lead to misinterpretation and hence better interpretation of mammograms is of higher importance. Recent work in advanced mammography imaging frameworks have paved the way for better diagnosis of abnormalities present in the breast. Computer-Aided Diagnosis (CAD) system have been designed for analysis of mammograms to classify the input image of mammogram as normal or abnormal(benign, malignant). The images used in all of the work have been extracted from the MIAS database of mammograms.

In this thesis, the implementation and contributions aim at developing new CAD system using deep learning to solve the challenges present in breast cancer diagnosis. The new CAD is used such that the features can become more efficient for classification of the pattern of tissues. In this context, firstly an existing CAD system have been implemented for pre-processing, segmentation of the Region of Interest(ROI), extraction of features, and
classification of breast cancer into classes. In this scenario, machine learning techniques made a breakthrough with automating the process of mammogram classification but deep learning, a subset of machine learning, increased the accuracy of detecting breast cancer even more. So later, to use deep learning concept in this domain, transfer learning has been done on pre-trained models. And further, one novel architecture and a tool for mammogram classification have been developed. At last all the observations were analyzed.

Firstly an existing CAD system have been implemented. In this the original mammograms were preprocessed to remove any noise present. After that segmentation have been done to segment the portion of the image with suspicious parts from the preprocessed images. The features were extracted to classify them. The extracted features were fed into
different classifiers to classify them as normal, benign or malignant.

Deep learning models AlexNet and VGG-19, which are pretrained on ImageNet database were used as feature extractors. Initially, MIAS database was fed to them after preprocessing. The extracted features or activations from intermediate layers were then classified by using different classifiers to classify them as normal or abnormal. For AlexNet, a variety of adaptive histogram equalization called Contrast-Limited Adaptive Histogram Equalization (CLAHE) based dataset was used and its performance was compared with the dataset without CLAHE. As the dataset without CLAHE performed well, it was considered further to be used with VGG-19. At last, results from both the models have been compared.

Finally, a self-designed architecture have been proposed by using Batch Normalization and Regularization and implemented by using best practices of Software Engineering. And later it’s performance was compared with all other implemented methods and pre-trained models used. Best accuracy was given by proposed model. Later mammogram classification tool has been developed which will first load the model, which was obtained after saving the proposed model. This tool would predict whether input mammogram is abnormal or normal based on the training provided to proposed model on MIAS dataset. At last analysis of software engineering and deep learning in CAD system for mammogram classification was done.