![]() | Only 14 pages are availabe for public view |
Abstract Colorectal cancer (CRC) is considered one of the most deadly cancer types nowadays. It is rapidly increasing for many reasons such as the unhealthy lifestyles followed by many people, water and food pollution, and aging. Additionally, the recent advancements in medical diagnostic technology have also led to the early discovery of diseases which contributed to the identification of more cases. Accordingly, it is significant to detect CRC in the early stages as this can help stop its growth by providing the necessary treatments which in return saves many people’s lives. There are different types of medical tests that doctors can perform to diagnose CRC, however, biopsy using histopathological images is considered the “golden standard” for CRC diagnosis. Due to the advancements in technology and hardware resources, it has become feasible to make use of various deep learning techniques in building computer-aided diagnosis (CAD) systems that help doctors in the process of disease diagnosis and prognosis. These systems can process medical input samples and then make a decision regarding whether the input sample shows any symptoms of cancer. They can additionally be used to determine the cancer stage and where it exists with an acceptable degree of confidence when compared to the doctor’s diagnosis. This thesis studies the colorectal cancer classification problem using weakly annotated histopathology whole slide images (WSIs) based on deep learning techniques. Its contribution starts by introducing efficient preprocessing steps to be applied to the raw dataset which outputs an optimized version that is a computational power-efficient representation of the dataset. To measure the effect of these preprocessing steps from the performance perspective, different models and experiments were performed based on the multiple instance learning (MIL) algorithm which resulted in achieving an accuracy of 89.58% compared to 84.17% of the baseline work inaddition to the superiority in optimizing the learning and testing processes and choosing a less complex model size. Moreover, the thesis proposes a modification to the learning algorithm (MIL) by relaxing the MIL conditions and introducing three WSI-label prediction functions to be integrated with the inference process of MIL. This resulted in a noticeable improvement by achieving an accuracy of 93.05% compared to 84.17% of the baseline work, maintaining the superiority of the learning and testing processes and model simplicity. Additionally, the performed experiments using the weak annotations alone outperformed the baseline work evaluation results obtained by applying a pre-training step using a set of strongly annotated slides. The methodology introduced in this thesis can be applied to similar cancer problems which opens the door for detecting and diagnosing cancer earlier and saving more lives. |