Segmentation on PA chest x-ray images

Abstract

In the field of medical image processing, object detection techniques play a key role for Computer Assisted Diagnosis (CAD), as well as for feature extraction tasks for other algorithms. One such practical problem is the detection of lung nodules on PA chest X-ray images which can, for example, help to increase the detection of lung cancer at an early stage. In this paper our aim is to compare the performance of different Convolutional Neural Network architectures, such as simple feed-forward networks and their combination with YOLO V1’s head and UNET-s combined with Densely Convolutional (DENSE) blocks on this problem. Furthermore, we provide insight into techniques used for fitting these networks on smaller datasets, by training and testing our solutions on the JSRT dataset, which only consists of 247 images. While we don’t always manage to achieve a good fit, by utilizing the proposed augmentation and preprocessing techniques, we manage to substantially decrease the loss on the validation dataset, as well as get qualitatively better results. Finally, Xray images are often only provided in an unnormalized DICOM format, where the choice of the utilized normalization method of the input images often becomes crucial in regard of the performance of neural networks. For this task we also analyze multiple methods, such as min-max normalization with the possibility of detecting outlier intensities, histogram equalization, and normalization by creating a simple, rough segmentation of the lung through traditional image processing methods and observing intensities in that area.