AI Reads MRI Scans to Reveal Hidden Survival Clue in Liver Metastases

When colorectal cancer spreads to the liver, the stakes could hardly be higher. Patients typically receive chemotherapy before surgery to shrink these metastatic tumors, a strategy known as neoadjuvant chemotherapy, and oncologists have long relied on how much tumor tissue dies off—captured by a pathology measure called tumor regression grade, or TRG—to gauge how well the treatment worked and how the patient will fare. The problem is that TRG can only be assessed definitively after the tumors are removed and examined under a microscope. By then, the treatment decisions have already been made. A new study published in BMC Medical Imaging by researchers at Peking University Cancer Hospital suggests that artificial intelligence applied to routine liver MRI scans may be able to peek inside that pathology report before the operation, and, more strikingly, that the mix of treatment responses scattered across a patient’s liver lesions carries powerful prognostic information that no single summary number can capture.

The research team, led by Qian Xing, Yong Cui, Xiao-Lei Gu and senior author Ying-Shi Sun, set out to do two things. First, they built a radiomics model—a machine learning pipeline that extracts hundreds of quantitative features from medical images, such as texture, intensity patterns and shape descriptors invisible to the human eye—to predict the tumor regression grade of individual liver metastases from gadoxetic acid-enhanced MRI. Gadoxetic acid is a liver-specific contrast agent that is taken up by functioning hepatocytes, allowing radiologists to image metastases with unusual clarity during the hepatobiliary phase of the scan. Second, and arguably more innovatively, they asked whether the heterogeneity of predicted responses across all of a patient’s lesions—some shrinking dramatically while others barely respond—could itself serve as a warning sign for future recurrence.

The study was a single-center retrospective analysis that included 295 liver lesions from 83 patients with colorectal liver metastases, split into a training group of 199 lesions and a validation group of 96 lesions. For each lesion, the researchers drew regions of interest on the latest preoperative MRI and computed radiomic features, ultimately screening down to 25 informative features that fed the predictive model. Rather than using the conventional five-tier TRG scale directly as the training label, the team used tumor residual rate—the percentage of original tumor tissue still alive after chemotherapy—as a continuous measure of response. This choice allowed the model to learn a graded, quantitative relationship between image texture and biological response rather than forcing predictions into coarse categories.

How well did the model perform? The correlation between the radiomics score and the actual tumor residual rate was moderate in both the training and validation groups, with coefficients of determination of 0.360 and 0.440 respectively. In plain terms, the AI could explain roughly a third to nearly half of the variation in how much tumor tissue survived chemotherapy, based purely on the appearance of the lesions on MRI. The authors describe this as modest predictive performance, and they are careful not to overclaim. Radiomics is not yet a substitute for pathology. But the fact that a noninvasive scan can capture a meaningful fraction of treatment response information at all is significant, because it opens the door to assessing response while the patient is still on chemotherapy, when the treatment plan can still be adjusted.

The truly headline-grabbing finding, however, concerns heterogeneity. The researchers defined radiomics-predicted TRG heterogeneity, abbreviated R_H, as the situation in which a single patient harbors both good predicted responses—predicted TRG 1 to 3, corresponding to a predicted tumor residual rate of 50 percent or less—and poor predicted responses—predicted TRG 4 to 5, with a predicted residual rate above 50 percent—across different liver lesions. This is the imaging equivalent of what pathologists have long observed under the microscope: metastases in the same liver do not respond uniformly to the same drugs. Some lesions melt away while others stubbornly persist, and this mosaic of sensitivity reflects the underlying biological diversity of the cancer.

To test whether this imaging-detected mosaic matters clinically, the team tracked local tumor disease-free survival, or LTDFS, a measure of how long patients remain free of locally recurring disease in the liver. Using Kaplan-Meier survival analysis, they compared the 13 patients whose lesions showed heterogeneity (R_H positive) with the 70 patients whose lesions were uniformly predicted to respond (R_H negative). The difference was stark. Patients with heterogeneous predicted responses had a median local tumor disease-free survival of just 3.0 months, while those without heterogeneity enjoyed a median of 10.2 months—a more than threefold difference that reached statistical significance with a p value of 0.002.

The researchers then pushed further, using Cox regression analysis to determine which factors independently predicted LTDFS when considered alongside clinical characteristics. Three variables emerged as independent predictors: the presence of radiomics-predicted TRG heterogeneity (p = 0.012), the best radiomics-predicted response across all lesions, termed R_min (p = 0.020), and whether the patient had undergone intraoperative radiofrequency ablation, a technique that destroys tumors with heat (p = 0.010). The appearance of R_min among the independent predictors is particularly intriguing. It suggests that the single best-responding lesion in a patient’s liver may carry more prognostic weight than the worst one, hinting that a cancer capable of a strong response somewhere may have a fundamentally more favorable biology than one that responds poorly everywhere.

Why should heterogeneity be so ominous? Tumor heterogeneity is one of the central challenges in modern oncology. A cancer that presents a mixed face to chemotherapy is likely to harbor resistant clones that survive treatment and seed recurrence. When those clones are distributed across multiple lesions in the liver, even complete surgical resection of the visible disease may leave behind microscopic resistant cells destined to regrow. Conventional imaging assessments, such as the RECIST criteria that track changes in lesion size, are poorly equipped to capture this biological diversity, because size is a crude proxy for response. A lesion can shrink substantially yet still be packed with viable tumor, while another may be replaced almost entirely by scar tissue. Radiomics, by quantifying internal texture and signal patterns, offers a window into these differences that simple measurements cannot provide.

The study has limitations that the authors and readers should keep in view. It was retrospective and conducted at a single center, so the model will need external validation in independent cohorts before it can inform clinical decisions. The number of patients with heterogeneity was small—13 individuals—which, while sufficient for statistical significance, leaves room for uncertainty around the precise effect size. The predictive performance of the radiomics model itself was modest, meaning that refinement of feature selection, segmentation and modeling approaches will be needed. Nevertheless, the prognostic signal from heterogeneity was strong enough to stand out even with a first-generation model, which bodes well for improved versions.

The broader implications are considerable. If validated prospectively, a radiomics-based assessment of response heterogeneity could be performed on the same gadoxetic acid-enhanced MRI scans that patients already receive before liver surgery, adding no additional imaging burden. Surgeons could use it to decide which lesions merit resection versus ablation, and oncologists could identify patients whose mixed responses warrant intensified systemic therapy or closer surveillance. The study also adds to a growing body of evidence that artificial intelligence can extract clinically meaningful information from images that radiologists cannot see with the naked eye. For patients with colorectal liver metastases—a disease where the liver remains the dominant site of treatment failure and where every month of disease-free survival matters—a smarter read of a routine scan could ultimately help tailor therapy to the true, uneven biology of each person’s cancer.

Subject of Research: Radiomics-based prediction of tumor regression grade heterogeneity on MRI for prognostic assessment in colorectal cancer liver metastases after neoadjuvant chemotherapy

Article Title: Radiomics-predicted TRG heterogeneity based on gadoxetic acid-enhanced MRI and its prognostic value in patients with CRLM after neoadjuvant chemotherapy

Article References: Xing, Q., Cui, Y., Gu, X.-L., Zhu, H.-T., Li, X.-T., & Sun, Y.-S. (2026). Radiomics-predicted TRG heterogeneity based on gadoxetic acid-enhanced MRI and its prognostic value in patients with CRLM after neoadjuvant chemotherapy. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02879-y

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02879-y

Keywords: radiomics, colorectal cancer, liver metastases, gadoxetic acid MRI, tumor regression grade, heterogeneity, neoadjuvant chemotherapy, machine learning, prognosis, disease-free survival, medical imaging, artificial intelligence