Convolutional Neural Networks for Medical Image Classification

Abstract

Medicine is one of the most vital areas of life, and it has shown true potential for the implementation of Artificial Intelligence tools, including patient management, diagnostics and cancer research. Within medicine, gynecology is a noticeable gap for development, due to the scarcity of personnel, lack of specialized education and negligence. As a result of these obstacles, there is a delay in the diagnosis of many serious conditions, and they only get proper attention once they turn cancerous. The noncancerous conditions are just as serious, as women report years lost to disability to them. The purpose of the thesis was to develop a deep learning, convolutional neural network based classification program that helps with the acceleration of diagnosis of gynecological conditions, which currently sets a noticeable gap within medicine. Upon utilization of the prototype, the examination time can be accelerated, alongside the elimination of countless control visits before the actual diagnosis of the underlying condition. The final working prototype can carry out a classification exercise upon ultrasound images fed into it by the user. Several research questions were driving this paper forward, including what the right CNN models are for medical image classification, what steps of data preparation are needed for the ultrasound images, what the right hyperparameter-tuning method is for medical image classification and how this program can be implemented into medical systems? This practical thesis was based on a personal project, where a multi-model classification program was developed from ResNet50, MobileNetV2 and GoogLeNet. These models were imported via transfer learning from Python’s machine learning library, PyTorch, then ensembled for a final classification exercise. Open-source datasets were utilized, which included ultrasound pictures of ovarian conditions, with multiple classes: clean, simple cyst, chocolate cyst, polycystic ovarian syndrome, teratoma, high grade serous cystadenocarcinoma. The project is based on said ultrasound images, but is also capable of being retrained on other medical imaging techniques, such as MRI or CT. The project is a deployable, scalable Python software package, which can be easily integrated into already existing medical viewing software or a standalone web application. Security, data privacy and sustainability aspects were also discussed in this paper. Based on the research tied to this project, and the success of the prototype, the utilization of such a tool is suggested. By this, the diagnosis based on medical imaging can be accelerated, while also directing attention to the obstacles that modern gynecology still faces.