A Master of Science thesis in Electrical Engineering by Mahmoud Rezk entitled, “Combining Saliency with Prediction for Endoscopic Diagnosis”, submitted in May 2020. Thesis advisors are Dr. Usman Tariq, Dr Abhinav Dhall and Dr Hasan Al Nashash. Soft copy is available (Thesis, Approval Signatures, Completion Certificate, and AUS Archives Consent Form).
Healthcare sector has advanced tremendously in the past few years. With the advancement in technology, many image diagnostic techniques have been introduced to help doctors in identifying diseases and abnormalities inside the human body. However, the increase in population and access to affordable healthcare have increased the patient population significantly, which requires a bigger infrastructure in medical diagnostics. The demand and supply imbalance of expert doctors in the field had led to the increase in healthcare bills. As a solution to the scarcity problem, one of the advancements that has been introduced to this sector is automated diagnostics using artificial intelligence (AI). The automated systems are made to help doctors in two ways. Firstly, they decrease the time required by the doctor to diagnose the patient and, secondly, they act as a second layer of diagnostic verification. This thesis aims to automate the classification of endoscopic images to eight disease and non-disease classes using a deep network architecture that would detect the salient region and classify the images accordingly. This thesis further studies the effect of jointly performing both tasks on the overall quality of attention masks and the classification results. The automated system is achieved by concatenating a U-net architecture to a dense-net architecture to jointly predict the salient medical masks and classify them to their respective classes. Furthermore, the automated system proved that medical image masks can be achieved by transfer learning the knowledge learned from natural images. Additionally, jointly predicting the masks and reusing the masks for classification demonstrated that the joint behavior would increase the classification accuracy.