Making diagnostics from medical images is a routine process. The digital tomosynthesis is a new method in examining the chest (lungs) by x-ray scans. During this process, tens of x-ray scans (40-60) are captured of a patient, in a limited impact angle. Based on these x-ray scans, coronal slice images of the chest are provided by using a reconstruction process which is quite similar to the method used at CT scans.
The sensitivity and diagnostics value of these slice images are much higher than the ones of the ordinary x-ray scans, although the emitted radiation dose is much lower compared to CT scans, and just a few times more than in case of ordinary x-ray scans. In medical control scans we would like to detect abnormalities (e.g.: TBC, malicious nodules) in the reconstructed images.
According to recent publications , pulmonologists are able to detect 51% of the early stage lung tumors in DTS reconstructions, while after interpreting the CT scans of the examined patients, 91% of these tumors were localized in DTS scans.
The existing digital image processing methods would provide a great improvement in assisting these tasks. However based on my recent knowledge, there is no such solution on the market which aims the DTS CAD (Computer Aided Detection).
In this thesis, I designed and implemented a prototype of a computer based image processor application. In order to perform this, I used deep neural networks. For building up, and validating the system, I had DTS scans of 8 patients, and an additional scan of an artificial lung “phantom”. I used Tensorflow framework to implement the neural network. Furthermore, I used convolutional neural networks for image processing. I used a method called transfer learning for training my neural network. With this method, I achieved 90% sensitivity, with 15 false positive detections per image in average. Due to the number of the available DTS scans, the results of the validation of the implemented CAD system should be treated with caution.