Polyp detection and tracking with advanced deep learning methods

Colorectal cancer is initially seen as a polyp in the mucus layer of the large intestine and, if left untreated, turns into cancer over time. Colorectal cancer, which has attracted attention with the number of cases close to one million in recent years, is shown among the deadly and aggressive cancer types. Although colon polyps are benign at first, they can turn into malignant cancer over time if they are not diagnosed and treated early. In this thesis, it is aimed to make automatic detection and follow-up of large intestine polyps. The problem of automatic detection of polyps by using deep learning methods and tracking them with object tracking algorithms is emphasized. Faster R-CNN model, which is one of the deep learning methods that has come to the fore in recent years, has been used for polyp detection. Object detection model training was carried out with Tensorflow Object Detection API. The LDPolypVideo dataset, which is 4 times larger than the new and existing datasets, was used in the training of the object detection model. Kalman filter, which is one of the object tracking algorithms, was used for polyp tracking. The object detection model we trained achieved an acceptable success with an mAP value of 86.37%. In tests performed with the trained object detection model, it was observed that our model detected polyps and marked them with bounding boxes. It has been seen that the object tracking method applied together with the object detection can be followed by labeling the polyps in the future. In this thesis, automatic polyp detection was performed using deep learning methods in a new colonoscopy dataset. The results obtained serve as an example to the studies in this field.