Detection of Prostate Cancer Using Deep Learning Techniques and Positron Emission Tomography/Computed Tomography (PET/CT) Images رسالة ماجستير

Abstract

Prostate cancer represents one of the most frequent malignancies in men globally, and accurate lesion detection in medical imaging is very important for early diagnosis and treatment. Modern imaging modalities such as Positron Emission Tomography/Computed Tomography (PET/CT) imaging yield very valuable information for diagnostics. At the same time, the process of lesion identification in these images is very time-consuming and not very accurate. The proposed thesis investigates the potential of the use of object detection models based on deep learning for the automatic detection of lesions in the prostate. Several state-of-the-art object detection models, such as YOLO-based object detection models and Roboflow OpenYOLO, were trained and compared based on the same data set and evaluation procedure. The Models were trained and evaluated using a local dataset collected from Augusta Victoria Hospital (AVH) in Jerusalem, comprising PET/CT images from 200 prostate cancer patients, totaling approximately 2,735 images focused on the prostate region, to my knowledge, PET/CT imaging has not been used in prior local studies, making this dataset a novel resource. Images were annotated with the help of expert radiologists. This dataset demonstrates the applicability of transformer-based detection models to real-world clinical data. Because of its high accuracy and robustness, RF-DETR was chosen as the final model for further testing. The chosen model was initialized with weights that were pretrained using COCO, and the usual COCO metrics, such as precision, recall, and Average Precision (mAP), were employed for evaluation. The experimental results demonstrate that the fine-tuned RF-DETR has achieved a precision of 93.9%, recall of 92.6%, and mAP@50 of 93.4%. These measurements ensure high detection capability and accurate lesion location in a moderate overlap condition. Lower efficiency at higher intersection over union (IoU) thresholds, indicated by lower mAP@50-95 values, indicate the inherent challenge in precisely locating lesions at the boundary. The qualitative analysis has also substantiated these results with proper lesion location in both internal and external test sets, efficient processing of cases with no lesions, and successful detection of smaller lesions with challenging conditions. A detailed analysis involving true positive, false positive, and false negative values has also given an insight into the experiment results. In conclusion, this work has shown that a transformer-based models, especially RF-DETR, is an efficient tool for automatically detecting a prostate lesion from a medical image. Although there is still a problem in localization accuracy at a higher threshold of IoU, this study has a great potential for improvement and provides a good basis for further advancement by improving localization accuracy and multi-modal data integration.