Liver malignancies are commonly evaluated using contrast-enhanced computed tomography (CE-CT), but in high-volume, real-world radiology workflows missed or delayed diagnoses remain a clinically important problem. The authors frame a need for scalable diagnostic safety-net approaches that can be integrated into routine practice to reduce diagnostic errors and expedite clinical intervention.
The Liver DiagnOsis Network (LiON) is a CE-CT–based artificial intelligence system developed to support flexible multiphase image processing, to incorporate clinical data, and to operate in a workflow-compatible manner within existing radiology practice. The system was engineered to perform focal liver lesion detection and malignancy diagnosis and to provide outputs that can function as an additional AI reader alongside human radiologists.
Key reported capabilities include multiphase processing and integration of clinical metadata to inform diagnosis. The study notes that some components rely on proprietary internal infrastructure (for example, phase classification, image registration and rendering), and that certain implementation details are protected by pending patent applications.
LiON was trained on imaging and clinical data from 6,443 patients. Retrospective validation was performed across multicenter and real-world cohorts totaling 22,251 patients. In these retrospective evaluations LiON achieved a high diagnostic accuracy for liver malignancy, with an overall AUC of 0.975 (95% confidence interval (CI): 0.971–0.979). The publication presents multicenter performance figures to illustrate diagnostic discrimination across cohorts.
The authors emphasize both large scale and real-world diversity in the retrospective datasets used for validation, suggesting broad applicability across different clinical settings represented in their cohorts.
LiON’s performance was assessed in clinically relevant subgroups. In patients with hepatic steatosis, LiON achieved an AUC of 0.971 (95% CI: 0.952–0.985). In patients with cirrhosis, LiON achieved an AUC of 0.924 (95% CI: 0.901–0.946). These subgroup results indicate preservation of diagnostic performance in common liver disease contexts that can complicate imaging interpretation.
The investigators conducted a single-arm trial in routine clinical practice that enrolled 10,333 patients. In the trial LiON was deployed as an additional AI reader integrated within the existing clinical workflow rather than replacing human reading.
The predefined primary endpoint required that the lower bound of the 95% CI for malignancy-diagnosis AUC exceed 0.900. The trial met this primary endpoint, with LiON achieving an AUC of 0.952 (95% CI: 0.942–0.961) in the prospective deployment setting.
Secondary outcomes from the single-arm trial documented instances where AI–human collaboration changed clinical reporting and management. Specifically, the combined workflow identified 51 previously overlooked lesions, of which 15 were malignant. These findings prompted 37 amended radiology reports and 22 escalations to multidisciplinary teams, resulting in clinical management changes for a subset of patients.
The authors interpret these secondary outcomes as evidence that a workflow-compatible AI support system can help detect missed findings and trigger downstream clinical actions. They note that these observations suggest potential for reducing missed or delayed diagnoses, though they stop short of claiming demonstrated improvements in hard clinical endpoints.
Patient-level datasets analyzed in the study are not publicly available because of ethical restrictions and patient confidentiality protections. The authors state that qualified researchers can request access to anonymized individual-level data and related clinical documentation by contacting the corresponding authors. Access requests will be evaluated by institutional data governance committees, and may require institutional review board approval and a data use agreement. The authors indicate requests will be processed within a maximum of 6 weeks.
Full LiON source code is not publicly released owing to dependencies on proprietary infrastructure and pending patent protection (CN116993663A). However, the report provides experimental and implementation details in Methods and Supplementary Information to support replication using non-proprietary libraries. The authors made the key algorithmic components and inference logic for focal liver lesion detection and diagnosis available in an open-source repository (https://github.com/alibaba-damo-academy/pixel-lesion-patient-network) and implemented the framework in PyTorch.
The authors acknowledge that, while LiON showed high discrimination in retrospective and single-arm prospective evaluation, further evidence is needed to assess effects on clinical outcomes. They recommend prospective comparative studies across diverse healthcare systems to determine whether AI deployment improves patient outcomes and to evaluate generalizability beyond the cohorts studied. The study also notes constraints on data and code sharing due to confidentiality, proprietary infrastructure and intellectual property considerations.
Overall, the reported work describes a large-scale development and validation effort for an AI system intended to function as a diagnostic safety net in CE-CT assessment of focal liver lesions, with evidence of strong diagnostic performance and examples of real-world clinical impact during a single-arm trial. The authors stop short of claiming definitive outcome benefits and call for additional prospective comparative research.