Detection-guided kidney tumor segmentation in 3d from abdominal computed tomography images

Abstract

Kidney tumors are one of the predominant causes of kidney cancer, posing a significant threat to human health. Early detection plays a crucial role in improving treatment outcomes for patients. Radiology imaging techniques such as ultrasound, MRI, CT, and X-ray are employed for kidney tumor detection. However, specialists have to manually locate the tumors and cysts from CT scans and with the high incidence rate of kidney cancers, this time-consuming approach puts pressure, leading to increased medical errors. To improve the diagnostic procedure, this thesis study proposes an automated detection-guided segmentation method. The two-step approach uses the YOLOv8 object detection model to detect the kidney region and the U-Net architecture to segment the abnormal region composed of tumor, cyst, or both. This method is evaluated in a 3D subject-wise evaluation technique for abnormal region segmentation. Four comparative methods are studied. Among these four, two are baseline methods, baseline U-Net and baseline YOLO, which are compared with the proposed method to show that guided frameworks perform better than single task frameworks. Two segmentation-guided methods, segmentation-guided classification and segmentation-guided segmentation, are compared to show how proposed method outperforms in the more relevant metric of recall, making it more suited to be a diagnostic method for detecting kidney abnormal regions. The results show that for all IoU thresholds ranging from 10% to 50%, the average recall for the proposed method is 1.000 across all thresholds. When comparing the approaches of the proposed method and the comparative methods, the proposed detection-guided segmentation method proves to be the most suitable as an automated method for kidney tumor diagnosis.