AI System Matches Radiologist Accuracy in Liver Cancer Detection Without Contrast Dye

A new artificial intelligence system has demonstrated it can accurately diagnose liver lesions using only non-contrast MRI scans, potentially eliminating the need for gadolinium-based contrast agents that carry risks for certain patients, according to a multicenter study published in npj Digital Medicine.

The system, called PRISM (Plain MRI Recognition and Interpretable System for hepatic Malignancy), was developed by researchers from nine institutions in China and tested on 12,823 patients. It achieved 98.1% accuracy in distinguishing benign from malignant liver lesions and 85.9% accuracy in classifying four malignant subtypes—all without requiring contrast enhancement.

Boosting Radiologist Performance While Saving Time

In a multireader evaluation involving 13 radiologists of varying experience levels, PRISM demonstrated significant benefits when used as a diagnostic aid. Radiologists using AI assistance improved their benign-malignant diagnostic accuracy by 4.4% and their four-subtype classification accuracy by 12.1% compared to unaided interpretation.

The time savings were substantial. AI-assisted interpretation reduced the time needed for benign-malignant classification by 30.5% (10.9 seconds) and slashed subtype classification time by 66.6% (61.8 seconds). These efficiency gains could prove valuable in clinical settings facing workflow bottlenecks or limited subspecialty expertise.

Three-Stage Diagnostic Approach

PRISM employs a three-stage diagnostic paradigm: automated detection and segmentation of focal liver lesions, benign-malignant classification, and finally four-subtype malignant classification. This structured approach mirrors clinical reasoning while maintaining interpretability—a key consideration for regulatory approval and clinical adoption.

The system's performance was validated across five test cohorts from multiple institutions, demonstrating generalizability beyond a single center. A prospective validation study of 1,147 consecutive patients further confirmed real-world applicability.

Triage Capability for Low-Risk Patients

One of PRISM's most clinically relevant features is its ability to identify low-risk patients who may not require immediate specialist referral or invasive procedures. The system flagged 79.3% of patients as low-risk with a 99.1% negative predictive value, potentially streamlining care pathways and reducing unnecessary interventions.

For patients, the contrast-free approach offers several advantages: elimination of gadolinium exposure (particularly relevant for those with renal impairment or concerns about gadolinium retention), reduced examination time and cost, and potentially faster diagnosis through automated triage.

Implications for Clinical Practice

The study positions non-contrast MRI combined with AI as a viable first-line diagnostic pathway for focal liver lesions, potentially disrupting the standard contrast-enhanced MRI workflow. This could prove especially valuable in resource-limited settings, for patients requiring serial monitoring, or in populations where contrast administration carries elevated risks.

The research was supported by the National Natural Science Foundation of China, Shanghai Sailing Program, Shanghai Municipal Health Commission, Shandong Provincial Natural Science Foundation, and institutional grants, with Ruijin Hospital serving as the lead institution.

The system was developed and tested primarily in Chinese institutions, and further validation in diverse international populations would strengthen evidence for global adoption.

Sources

Interpretable deep learning for three-stage focal liver lesion diagnosis on non-contrast MRI: a multicenter, prospective study — npj Digital Medicine