Aim of the project was the development of a computer-aided design algorithm (CAD) that identifies possible types of thoracic diseases in chest X-rays. Radiologists can therefore be supported in the X-ray based diagnostics of cancer and other diseases. The developed model highlights suspicious areas in the X-ray images and provides a classification for chest specific disease types.
The project was based on a training dataset of 100,000 anonymized chest X-Ray images together with the corresponding diagnosis type.
Challenges & Solutions
At first, the X-Ray images were pre-processed, so that color differences were coherent within the set of X-Ray images and the areas of interest were identifiable.
Then feature processing was applied. Using algorithms the main features in the x-ray images clustered image areas into a set of different color ranges. These color clusters highlighted and captured important areas of interest.
Based on the extracted features, the relevant characteristics (such as size, orientation, shape and location) of the different diagnostic types were selected.
Having extracted and selected the important features of the X-Ray images, Machine and Deep Learning algorithms could be trained to detect areas of interest in X-Rays and classify them to one of the diagnosis types according to the selected features.
The Deep Learning and Image Processing Framework was developed and successfully applied in the X-ray data set, ready to support radiologists in identifying suspicious areas within thorax X-rays and support decision making in diagnostics.