A magnetic resonance imaging approach has shown potential to estimate tumour grade before treatment in intrahepatic mass-forming cholangiocarcinoma and provide preliminary information about survival risk. The retrospective multicentre study, published in Insights into Imaging, used preoperative scans from 333 patients treated at three hospitals in China. The framework combined deep learning across image slices and whole tumour volumes with radiomics, clinical, laboratory and conventional imaging information. External validation favoured the integrated model over models using radiomics or conventional variables alone, while its improvement over deep learning alone was limited. Survival-related findings were exploratory because follow-up information was available for only part of the cohort.
Must Read: PSMA PET/MRI Refines LR-3 Liver Assessment
Building a Multidimensional MRI Framework
Intrahepatic mass-forming cholangiocarcinoma is the most common form of intrahepatic cholangiocarcinoma and is associated with rapid growth, early spread and poor outcomes. Histological grade can inform treatment planning, including decisions concerning surgery and treatment before or after surgery. Biopsy remains the basis for tissue assessment, but sampling error can limit grading accuracy in heterogeneous or mucin-rich tumours. The MRI framework was therefore designed to provide complementary non-invasive information before treatment rather than replace histopathological assessment.
The cohort included patients with confirmed disease who underwent multiparametric MRI before surgery or biopsy between 2018 and 2024. Two hospitals contributed the development cohort and a third supplied an independent validation group. Tumours were classified as low grade or high grade according to World Health Organization criteria. Diffusion-weighted and T2-weighted MRI were used because they provide information about tissue cellularity, extracellular composition and tumour structure.
Radiologists assessed features such as tumour size, margins, vascular involvement, lymph nodes and signal patterns. Tumours were manually outlined on both MRI sequences and reviewed by an experienced radiologist. Agreement between radiologists was high and differing assessments were resolved through joint review. Tumour size, margin characteristics, vascular involvement and a target-like pattern on T2-weighted imaging remained associated with grade after statistical assessment. Radiomics processing also selected a smaller group of quantitative image features from several thousand initial measurements.
Integrated Model Leads External Validation
The deep learning component combined two image-analysis approaches. One processed neighbouring slices to capture local tumour heterogeneity, while the other assessed the tumour as a complete three-dimensional volume. Their outputs were merged to combine fine image detail with broader spatial information from diffusion-weighted and T2-weighted scans.
Four prediction models were compared. One used clinical, laboratory and conventional imaging variables, another used radiomics, a third used the combined deep learning output and the final model integrated all available sources. In the independent cohort, the deep learning model performed better than either image architecture used alone. The integrated model achieved the strongest overall discrimination, with an area under the receiver operating characteristic curve of 0.843.
The integrated model also showed stronger performance than the radiomics and conventional-variable models. Its improvement over deep learning alone was modest and did not reach statistical significance, suggesting that most predictive information came from the MRI-based deep learning features. Decision analysis indicated that the integrated model offered the greatest benefit across a broad range of prediction thresholds.
Interpretability methods were used to show which image regions and variables influenced predictions. Attention maps focused on diffusion-weighted regions linked with dense tumour tissue and T2-weighted areas associated with necrotic or mucin-rich components. Feature analysis showed that deep learning contributed most to the model output, followed by clinical, laboratory and imaging variables, while radiomics made a smaller contribution.
Prognostic Findings Require Further Validation
The integrated model divided patients into predicted high-risk and low-risk groups using a threshold established in the development cohort and retained for external validation. Overall survival information was available for 175 patients. Across the available follow-up data, the predicted high-risk group had shorter survival than the low-risk group and the same pattern appeared in both the development and validation cohorts.
These results remain preliminary. Survival data were missing for many patients because of loss to follow-up, recent diagnosis or incomplete records. Although the included and excluded groups did not differ significantly in their recorded baseline characteristics, the restricted sample limits confidence in the prognostic assessment.
Performance was also lower in external validation than during model development, particularly for identifying high-grade tumours. The participating hospitals differed in patient characteristics, laboratory findings, scanner settings and imaging protocols. Such variation provided a more demanding assessment of generalisability but may have reduced model stability. The external cohort was also relatively small, limiting the precision of the estimates.
Other limitations include the retrospective design, possible selection bias and manual tumour segmentation. Predictions may be less reliable when lesions contain extensive necrosis, haemorrhage or mucin, which can create ambiguous MRI signals. The evidence does not address routine workflow integration, cost-effectiveness or direct effects on clinical decisions. The framework therefore remains a potential decision-support method that should be considered alongside biopsy findings and multidisciplinary judgement.
A multidimensional MRI framework combining slice-based and whole-tumour deep learning with selected radiomics and clinical information supported preoperative grading of intrahepatic mass-forming cholangiocarcinoma. External validation favoured the integrated model over radiomics and conventional-variable approaches, although its added value beyond deep learning alone was limited. Interpretability methods linked predictions to relevant tumour regions and showed that deep learning features drove most model decisions. The framework also separated groups with different survival outcomes, but incomplete follow-up and retrospective data make this finding exploratory. Larger prospective multicentre evaluation is needed before routine clinical use.
Source: Insights into Imaging
Image Credit: iStock
References:
Zhuo L, Chen W, Song Z et al. (2026) Multidimensional deep learning for grading and prognostic assessment of intrahepatic mass-forming cholangiocarcinoma. Insights Imaging; 17, 187.
