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

This multicenter, retrospective study compiled data from 206 HCC patients with pathologically confirmed diagnoses, sourced from three distinct hospitals, ensuring a diverse and robust dataset for model training and validation. The investigative team focused sharply on the hepatobiliary phase images (HBP) of gadoxetic acid-enhanced MRI, a specialized imaging sequence known for its superior liver lesion characterization. To harness the full potential of this imaging modality, three variations of deep learning architectures were meticulously developed and assessed: two-dimensional (2D), three-dimensional (3D), and a novel 2.5-dimensional (2.5D) deep multi-instance learning (MIL) model.

Among these approaches, the 2.5D MIL technique emerged as the most potent, ingeniously integrating information from all axial slices encompassing the tumor and surrounding peritumoral regions. This comprehensive slice selection allowed the model to capture both intratumoral heterogeneity and critical peritumoral microenvironmental features, which are hypothesized to play pivotal roles in MVI development. Remarkably, this approach outperformed conventional models with area under the curve (AUC) values reaching 0.802 in internal validation and 0.759 in the external test cohort, highlighting its strong generalizability across patient populations.

Building on these promising findings, the researchers extended their methodology by incorporating additional MRI sequences—T1-weighted fat-suppressed (T1WI-FS) and T2-weighted fat-suppressed (T2WI-FS) images—into a multimodal prediction framework. The synergistic use of these complementary sequences aimed to further refine the predictive algorithm by capturing distinct tissue contrasts and pathological signatures of MVI. This multimodal deep learning model demonstrated exceptional predictive capacity, with AUC values soaring as high as 0.954 in the training set, and sustaining robust performance metrics with AUCs of 0.857 and 0.788 in independent validation and test sets respectively.

These advancements are not merely incremental; they represent a paradigm shift in medical imaging diagnostics for liver cancer. The exploitation of 2.5D MIL in this context signifies a novel computational strategy that leverages the spatial and contextual richness of MRI datasets more effectively than traditional machine learning or purely 2D/3D frameworks. By embracing the heterogeneous nature of tumor microenvironments across consecutive axial slices, this approach unlocks a more nuanced understanding of tumor biology, which is crucial for preoperative risk stratification.

Importantly, the ability to noninvasively predict MVI preoperatively holds profound clinical implications. Surgeons and oncologists can now potentially tailor treatment regimens with increased precision—deciding between surgical resection, transplantation, or adjuvant therapies based on individualized risk profiles. Moreover, early detection of MVI propensity may guide surveillance intensity post-operation, improving patient outcomes through timely interventions.

The retrospective design of this study, coupled with its multicenter data acquisition, lends credibility to the model’s robustness and applicability across different clinical settings. However, the authors underscore the necessity for prospective trials to validate and eventually integrate these computational tools into routine clinical workflows. Additionally, further research into integrating radiogenomics could uncover deeper mechanistic links between image features and genetic underpinnings of microvascular invasion.

On the technical front, the study exemplifies an exemplary application of advanced artificial intelligence methodologies in oncological imaging. The use of deep learning architectures capable of handling multi-instance learning tasks shows how algorithmic innovation can extract actionable insights from complex and high-dimensional medical images. Fine-tuning such models with diverse MRI sequences reinforces the importance of multimodality in capturing the multifaceted nature of cancer pathology.

This research also emphasizes the crucial role of peritumoral tissue analysis, an often-overlooked area in conventional imaging assessments. The findings suggest that changes in the tumor microenvironment surrounding hepatocellular carcinoma lesions carry significant predictive information, urging the clinical community to expand diagnostic focus beyond tumor margins. Such insights will likely spur new investigations into tumor-stroma interactions and their implications in cancer progression.

Future directions proposed by the research team involve the development of automated MRI preprocessing pipelines and user-friendly software interfaces to democratize the use of their predictive models in community hospitals and cancer centers worldwide. Integrating these outputs with electronic health records and decision-support systems could democratize access to personalized oncology care, aligning with global efforts toward precision medicine.

In conclusion, the creation and validation of a deep multi-instance learning model based on gadoxetic acid-enhanced MRI heralds a new age in the preoperative management of hepatocellular carcinoma. By harnessing the power of 2.5D imaging data and multimodal sequences, this approach offers an unprecedented window into tumor invasiveness, enabling clinicians to make better-informed decisions and ultimately improving patient prognoses. As artificial intelligence continues to evolve, such innovations underscore the transformative potential of combining computational prowess with clinical expertise in battling complex diseases like HCC.

Subject of Research: Microvascular invasion prediction in hepatocellular carcinoma using advanced deep learning models applied to gadoxetic acid-enhanced MRI.

Article Title: Deep multi-instance learning model based on gadoxetic acid-enhanced MRI for predicting microvascular invasion of hepatocellular carcinoma: a multicenter, retrospective study.

Article References: Luo, Y., Zhang, G., Zhong, S. et al. Deep multi-instance learning model based on gadoxetic acid-enhanced MRI for predicting microvascular invasion of hepatocellular carcinoma: a multicenter, retrospective study. BMC Cancer 25, 1626 (2025). https://doi.org/10.1186/s12885-025-14971-7

Image Credits: Scienmag.com

DOI: https://doi.org/10.1186/s12885-025-14971-7