Abstract:

According to the Association of American Medical Colleges, America’s shortage of radiologists and other physician specialists could surpass 35,000 by 2034. The situation is equally tight in the UK with just ‘2% of radiology departments able to fulfill their imaging reporting requirements within contracted hours’. Keeping in mind this lack of resources in medical imaging and the fact that Pakistan has the 10^th highest annual death rate for patients dying from lung diseases, there is a dire need for successful automation in medical imaging to ameliorate this situation. If not totally automated, Computer-aided diagnosis (CAD) can be at least interpreted as a second opinion in assisting the physicians during diagnosis. For diagnosis, Chest X-Rays (CXRs) are the cheapest and most commonly utilized tool used by radiologists for lung diseases. So, in this thesis, we are trying to understand how Radiologists diagnose different abnormalities using disease patches on CXRs. A lot of work is being done on the automated diagnosis, but the acceptability of these methods into routine clinical practice requires an explanation from the models as to why one should trust them. “Transparency, interpretability, and explainability are necessary to build patient and provider trust”. Understanding how a radiologist diagnoses and guiding the network to the specific disease areas can be advantageous for increasing its classification and localization ability which is what we are trying to show. This thesis involves the classification and localization of numerous lung diseases in CXRs. The dataset used is from a local hospital named Gulab Devi, comprising of around 500 images with corresponding disease annotation data of 233 patients collected at different time stamps. We propose using a pre-trained deep convolution neural network as a base model and attaching an attention-based mechanism to guide the network with the affected area to increase the model's performance and explainability. To incorporate the additional attention input, we use a simple and yet very effective bounding box-based attention strategy. It is inspired by approaches for conditioning probabilistic models on external inputs from the literature on generative image modeling. After the classification results, we employed a saliency mapping technique called GradCAM to extract the localization results from the model. Our proposed model trained on the Gulab Devi dataset achieves an AUROC classification score of 0.712 on five different lung diseases, improving 12.3% over the results of the state-of-the-art model on the Gulab Devi dataset. Qualitative results for the localization also show an improvement using our proposed model.

Evaluation Committee  

• Dr. Murtaza Taj (Supervisor)
• Dr. Muhammad Fareed Zaffar (Evaluator)

Zoom: https://lums-edu-pk.zoom.us/j/92808531431

 Meeting ID: 928 0853 1431