Deep learning algorithm enables lower limb venous thrombosis detection with CT venography

Lower limb computed tomography venography (CTV) has low success rates for deep vein thrombosis (DVT) diagnosis. This study applied deep learning to improve DVT identification.

Our study enrolled 119 positive DVT and 40 negative DVT, 111 positive pulmonary embolism (PE) patients and 20 negative PE. Two algorithms were evaluated: Faster R-CNN trained directly on CTV images, and YOLO11-nano pre-trained on computed tomographic pulmonary angiography (CTPA) then optimized on CTV via transfer learning. Three radiologists (3-5 years' experience) independently interpreted CTV images. Diagnostic performance was compared.

The YOLO11-nano model and Faster R-CNN achieved over 97% accuracy and sensitivity in diagnosing PE. Compared to the three radiologists in diagnosing DVT, YOLO also had an excellent accuracy (92.1%) and sensitivity (93.9%), superior to Faster R-CNN (66% accuracy, 68% sensitivity). The YOLO model exceeded 77% accuracy in pelvic and femoral-popliteal segments and 66.7% in detecting calf segment thrombi.

The CTPA-based transfer learning model significantly improved the feasibility of this method in routine CTV diagnostic performance, offering a promising approach for simultaneous PE and DVT detection.