| dc.description.abstract | Prostate cancer (PCa) is one of the most prevalent cancer types in the world, and it is the dominant type of cancer among males in Norway. Detection of PCa is done with the help of magnetic resonance imaging (MRI), which yields scans of a patient’s prostate. Radiologists analyze these scans to detect any potential lesions and to determine their severity. To confirm the presence of a lesion, the patient undergoes a prostate biopsy. Clinical analysis from radiologists are prone to subjectivity as a consequence of the challenging nature of detecting lesions. Recent developments in deep learning (DL) have shown potential to aid radiologists in detection and classification of lesions in the prostate, thus making their analysis more objective, and possibly reducing the amount of unnecessary biopsies. Nonetheless, the models applied in DL need large quantities of balanced and labeled data, which is often difficult to obtain.
This thesis delves into the methods and processes of semantically segmenting different types of lesions in MRI-scans of a patient’s prostate based on the lesion’s significance. Several approaches are used to reduce the complications caused by utilizing a small and imbalanced dataset, including the implementation of a fully convolutional encoder-decoder network with dropout, as well as removing MRI-slices where no lesions are present. To evaluate our methods, we use segmentation metrics such as Hausdorff distance (HD), dice similarity coefficient (DSC), and relative volume difference (RVD). Our best model is evaluated through bootstrapping, focusing on two cases: testing on the original dataset, and testing on the dataset where MRI-slices containing no lesions are removed. The resulting DSC for significant and insignificant lesions (gleason score ≥ 7 and gleason score < 7), were 29.00% and 22.44%, respectively. | |