Abstract

Although deep learning-based models show high performance in the medical field, they required large volumes of data which is problematic due to the protection of patient privacy and lack of publically available medical databases.
In this thesis, we address the problem of medical data scarcity by considering the task of pulmonary disease detection in chest X-Ray images using small volume datasets (<1000 samples). We implement three deep convolution neural networks pre-trained on the ImageNet dataset (VGG16, ResNet-50, and InveptionV3) and asses them in the lung disease classification tasks transfer learning approach. We created a pipeline that applied segmentation on Chest X-Ray images before classifying them and we compared the performance of our framework with the existing one. We demonstrated that pre-trained models and simple classifiers such as shallow neural networks can compete with the complex systems.
We also implemented activation maps for our system. The analysis of class activation maps shows that not only does the segmentation improve results in terms of accuracy but also focuses models on medically relevant areas of lungs.
We validated our techniques on the publicly available Shenzhen and Montgomery datasets and compared them to the currently available solutions. Our method was able to reach the same level of accuracy as the best performing models trained on the Montgomery dataset however, the advantage of our approach is a smaller number of trainable parameters. What is more, our InceptionV3 based model almost tied with the best performing solution on the Shenzhen dataset but as previously, it is computationally less expensive.