Abstract:
Spinal metastatic bone cancer is a severe complication of advanced cancer like lung, breast, and prostate, so its early diagnosis and correct classification should be crucial to enhance treatment planning, decision-making, and patient survival. Conventional methods of diagnosis based on manual assessment of MRI and CT are rather slow and subject to error, with the necessity to adopt the use of computers. In this research, a customized deep learning framework is proposed that integrates the You Only Look Once (YOLO) family of object detection models with a hierarchical fuzzy inference system to detect and classify spinal lesions in CT scans. The Cancer Imaging Archive (TCIA) was utilized as the source of the dataset of 1,367 CT images, which were provided as DICOM data, we carefully preprocessed the data and converted into PNG and annotated into three categories: Normal, Lytic and Blastic. Multiple YOLO versions (YOLOv8, YOLOv5, YOLOv9, YOLOv10, YOLOv11, and YOLOv12) were trained and evaluated under a uniform experimental setup, and performance was assessed using precision, recall, F1-score, and mean Average Precision (mAP). It was comparatively analyzed that YOLOv11 completed the stable results by providing a precision of 89.1% and recall of 85.8% in comparison to the previous versions of YOLO. And then, YOLOv11 was modified with Fuzzy logic based upon the need to consider uncertain situations and the use of rules and membership functions to modify the predictions assisted in turning the classifications more sure and easier to comprehend. The hybrid YOLO with Fuzzy system obtained a total accuracy of 93.69% that showed improved diagnostic performance over the baseline YOLO models. We applied the fuzzy logic as a post processing layer for handle uncertain result. The framework helps doctors make better decisions, reduces errors
