Why Ultrasound AI Matters Abdominal ultrasound is the frontline HCC surveillance tool globally because of its low cost and lack of radiation, but its diagnostic performance is heavily operator-dependent, with pooled sensitivity of 60-84% varying widely by sonographer experience. AI could standardize and improve these results.
The Study Scale This retrospective study from National Taiwan University Hospital is one of the largest ultrasound-based liver tumor AI studies to date: 1,576 patients, 4,599 ultrasound images, and 6,001 lesions spanning both malignant (HCC, cholangiocarcinoma, metastasis) and benign (cysts, hemangiomas, focal nodular hyperplasia) categories.
Two Parallel AI Tasks The study developed AI for two complementary tasks: (1) classification models distinguishing malignant from benign lesions and subtyping all eight lesion categories, and (2) a real-time detection model (YOLOR) that simultaneously localizes and classifies tumors within ultrasound frames.
Best Performance Among eight classification models tested, EfficientNet-B5 and Swin-B achieved the highest AUC of 0.90 for distinguishing malignant from benign hepatic lesions, while HCC specifically achieved the highest per-lesion classification accuracy among all malignant subtypes.
Dataset Composition The 6,001 lesions included 4,117 HCC (68.6%), 277 cholangiocarcinoma, 277 metastases, 301 cysts, 485 hemangiomas, 183 focal fatty sparing, 336 focal nodular hyperplasia, and 25 other benign findings. HCC dominance reflects the real-world composition of an Asian high-risk population with 48% HBV and 21% HCV prevalence.
Classification Models Compared Eight architectures spanning CNN and vision transformer approaches were trained: ResNet50, Xception, Inception ResNet V2, EfficientNet-B5, EfficientNetV2-S, EfficientNetV2-L, Swin-T, and Swin-B. All were trained with cross-entropy loss at learning rate 0.001, with gamma adjustment and augmentation (flip, rotation, crop) applied.
YOLOR for Real-Time Detection YOLOR (You Only Learn One Representation) was used for simultaneous lesion detection and classification. YOLOR's CSPDarknet53 backbone with SPP-PAN neck structure enables multi-scale feature extraction suitable for detecting lesions of varying sizes within a single ultrasound frame, outputting bounding box detections in real time.
Pathologic Ground Truth Malignant diagnoses were confirmed by CT, MRI, or pathology per AASLD and Taiwan Liver Cancer Association guidelines. Eight experienced ultrasonographers labeled each lesion with bounding boxes. Dataset was split 45:3:5 for training, validation, and testing, resulting in 1,061 training, 373 validation, and 142 test patients.
Top AUC Performance EfficientNet-B5 and Swin-B achieved the highest malignant/benign classification AUC of 0.90 in both validation and testing sets. ResNet50 and Inception ResNet V2 achieved AUC 0.88. All models showed some performance drop between validation and testing sets, consistent with test set difficulty.
HCC-Specific Accuracy Across all eight-class classification tasks, HCC achieved the highest per-lesion accuracy among malignant tumors in most models (e.g., 0.944 for Swin-T and 0.930 for EfficientNetV2-L in testing set), while cysts achieved the highest accuracy among benign lesions (1.000 in several models). Cholangiocarcinoma and metastasis had the lowest classification accuracy due to their rarity and morphological overlap.
YOLOR Real-Time Detection The real-time detection model (YOLOR) achieved mean Average Precision (mAP) scores of 0.5342 (YOLOR-W6) and 0.5631 (YOLOR-D6) for simultaneously detecting and differentiating malignant and benign lesions in the testing set - establishing feasibility for real-time AI-assisted ultrasound workflows.
Rare Lesion Challenge Cholangiocarcinoma (277 cases) and metastases (277 cases) showed significantly lower classification accuracy than HCC across all models, with some models achieving 0% accuracy for these classes. This reflects both the relative rarity of non-HCC malignant lesions in the training set and their greater morphological heterogeneity.
Addressing Operator Dependency The most immediate clinical application is standardizing ultrasound quality for less experienced operators. An AI assistant achieving AUC 0.90 could flag suspicious hepatic lesions in real time during the scan, prompting the sonographer to optimize image acquisition and capture additional views before concluding the examination.
Taiwan's HBV/HCV Context The study population (48% HBV, 21% HCV) mirrors the high-risk populations in Taiwan, China, Southeast Asia, and sub-Saharan Africa where HCC incidence is highest. Deploying AI-enhanced ultrasound in these regions could have the greatest global impact on HCC early detection rates.
Real-Time Decision Support YOLOR's real-time detection capability represents a step toward AI tools that provide live feedback during the ultrasound examination itself - fundamentally different from post-hoc image analysis. Real-time AI could guide the sonographer's probe positioning to ensure every concerning lesion is documented.
Scalability for Surveillance Programs Ultrasound surveillance programs for HCC in cirrhotic populations require consistent quality across many patients and many examiners over time. AI standardization of sonographic interpretation could enable high-quality HCC surveillance in community hospital settings without subspecialty hepatology expertise.
Single-Center Retrospective Design All images were acquired at National Taiwan University Hospital using a specific set of scanner models. Performance in community hospitals with different scanner brands, operators, and patient populations requires external validation before clinical deployment.
Static Image Analysis vs. Video The models were trained on still ultrasound images rather than video sequences. Real ultrasound examinations generate continuous video frames with motion artifacts and probe angle variation - conditions the current models were not specifically trained to handle.
Rare Malignancy Underperformance Cholangiocarcinoma and metastasis showed near-zero classification accuracy in some models, limiting clinical utility for the full spectrum of liver malignancies. Future studies should specifically oversample these rare but clinically important diagnoses.
YOLOR mAP Improvement Needed The YOLOR detection mAP of approximately 0.56 is promising but below the level needed for clinical deployment as a standalone tool. Improvements in lesion localization accuracy, particularly for small lesions, and testing in prospective real-time scanning settings are the next development priorities.