Pilot feasibility study of deep learning-based detection of malignant liver lesions in intraoperative ultrasound
Pablo Parra-Méndez1, Pablo Parra-Membrives2,3,4
1Software Engineer and Master in Artificial Intelligence, Loyola University.Seville, C/Rubi 35, 41927 Mairena del Aljarafe, Seville, Spain. pparra981@gmail.com.
Background:
Intraoperative ultrasound plays a central role in hepatobiliary surgery for the detection of liver lesions, but image interpretation is challenging and operator-dependent. Evidence supporting artificial intelligence-based detection tools in this setting remains limited. This study was designed as a preliminary pilot feasibility study of a future intraoperative support system for decision-making.
Methods:
A retrospective dataset of 1,035 intraoperative ultrasound images containing malignant liver lesions was collected and manually annotated by an experienced hepatobiliary surgeon. Images were divided into a training set (n = 935) and a validation set (n = 100), comprising 114 lesions. Because this was an exploratory pilot study, all malignant lesions were analyzed as a single detection class. Three detection strategies were evaluated: a two-stage convolutional architecture (Cascade R-CNN), a single-stage detector (YOLO11), and a transformer-based model (RF-DETR). Performance was assessed using sensitivity, precision, and mean average precision (mAP).
Results:
YOLO11 Medium achieved the highest sensitivity (63%), detecting 72 of 114 lesions while maintaining a precision of 62% and rapid offline inference. RF-DETR demonstrated the highest precision (80%) and superior localization accuracy (mAP@0.50 approximately 70%), with sensitivity close to 60%. Cascade R-CNN showed balanced but inferior performance compared with YOLO11 and RF-DETR.
Conclusion:
Automated detection of malignant liver lesions in intraoperative ultrasound appears feasible in the preliminary pilot study. However, the observed sensitivity remains insufficient for autonomous intraoperative decision-making. YOLO11 and RF-DETR showed complementary strengths in sensitivity and precision, respectively, and should be regarded as exploratory support tools requiring further training, patient-level validation, and external multicenter evaluation before clinical implementation.
