| Abstract [eng] |
This thesis investigates one-stage and two-stage approaches for kidney tumor segmentation in \ac{ct} images, when dealing with labeled data scarcity. The rapid development of \ac{dl} has turned medical image analysis into a research hotspot, but accurately segmenting tumors in medical images remains a challenging task, particularly when dealing with limited data, which is a common problem in the medical setting. The thesis proposes a novel approach to two-stage semantic segmentation, employing the latest YOLOv7 object detection model. Having implemented multiple approaches for dealing with labeled data scarcity, such as data augmentation and fine-tuning, the results of the experiments concluded the proposed fine-tuned two-stage approach to achieve an increase of 2.4\% overall Dice score, across all patients. Further investigation has found this method to be significantly more successful in the segmentation of a small tumor, which was undetected by the baseline one-stage approach. Although the results come in line with reviewed literature, they should be considered with caution, due to the poor population representability in the training and test set splits of a small dataset. |
