Depth of stromal invasion is a key factor in early-stage cervical cancer because it influences surgical planning and postoperative care. Conventional imaging already supports preoperative assessment, but diagnostic accuracy remains suboptimal. A 2026 analysis published in Insights into Imaging evaluated an explainable model combining MRI-derived imaging features, radiology-report information and clinical variables to assess invasion risk before surgery. The model links imaging patterns, report text and clinical data with machine learning and explainability tools, aiming to improve risk evaluation while keeping outputs interpretable for clinical use. The external validation cohort remained small.
Combining MRI, Reports and Clinical Data
The cohort included 319 patients with pathologically confirmed early-stage cervical cancer from two centres. Eligible patients had preoperative sagittal T2-weighted MRI within one month, confirmed squamous cell carcinoma and available clinicopathological information. Exclusion criteria covered missing imaging or radiology-report data, inadequate image quality, another malignancy and tumour volumes too limited for assessment. Institutional approval covered the retrospective design, with informed consent waived. MRI data supplied the imaging component.
A radiologist experienced in pelvic MRI manually marked tumour volumes on sagittal images. Image data were standardised before feature extraction, and reproducibility checks helped retain reliable imaging features. LIFEx software generated quantitative radiomic features in line with imaging biomarker standards. Radiology reports supplied the text component. These Chinese-language reports contained radiologist descriptions of imaging findings and differential diagnoses. Pre-processing removed wording that directly or indirectly indicated stromal invasion status, reducing the risk that the model would depend on explicit diagnostic conclusions. It also removed names and timestamps, excluded empty records and consolidated remaining content. A language-processing toolkit then converted report content into structured features for modelling and later comparison.
Must Read: Ultrasound Matches Advanced Imaging in Cervical Cancer
Building a Multimodal Prediction Model
The modelling process combined selected imaging, report-based and clinical information. The starting feature pool included several hundred MRI-derived features and several hundred text-derived features per patient. Reproducibility checks and statistical selection reduced this pool to a smaller group of radiomic and text features. The final feature set retained both BERT-derived text signals and MRI radiomic signals. Clinical variables were assessed separately, and age and tumour size emerged as independent risk factors for depth of stromal invasion. Pathological analysis separated patients into superficial stromal invasion and middle or deep stromal invasion groups. Several machine learning methods were tested across different model types.
Separate clinical, text and radiomic models were created first. Combined models then used report and MRI features, followed by report, MRI and clinical features together. Different algorithms performed best in different model groups. The best-performing models in the internal cohorts then moved into comparative analysis. Model evaluation used discrimination, calibration and decision-curve measures. Shapley Additive exPlanation added interpretability by showing how individual text, imaging and clinical features influenced predictions. The final comparison therefore covered performance, calibration and possible clinical usefulness, not only a single risk score.
Validation, Explanation and Limits
The combined text-radiomic-clinical model delivered the strongest overall performance. Its area under the curve reached 0.912 in training and remained high in internal and external validation. The text-radiomic model also performed strongly, while single-source models using only clinical data, report features or radiomic features showed lower performance. External validation included comparison with a standalone assessment by a senior radiologist, who reviewed the original MRI reports without access to pathological results or model predictions. The combined model showed higher performance than this standalone assessment in the external cohort. Calibration results showed close agreement between predicted probabilities and observed outcomes across cohorts. Decision-curve analysis indicated possible additional net benefit from adding clinical variables, despite no significant area-under-the-curve difference between the two fusion models.
Explainability analysis placed report-derived features among the most influential predictors, followed by radiomic features and clinical factors. The report-based signal still requires caution because indirect descriptors of tumour appearance may reflect part of the radiologist’s clinical impression. Limits remain important. The retrospective design may introduce selection bias. Only squamous cell carcinoma cases were included, very small tumours may be under-represented, and the external validation cohort was small. The external validation images came from a different vendor with distinct acquisition parameters, while the absence of explicit harmonisation and nested cross-validation may affect reproducibility.
The multimodal model combines MRI-derived features, processed radiology-report information and clinical variables to assess depth of stromal invasion before surgery in early-stage cervical cancer. Its strongest performance came from integrating all three data sources, with explainability analysis showing contributions from text, imaging and clinical features. The approach supports more structured preoperative risk assessment, but the results remain preliminary. Retrospective design, narrow histological inclusion, possible spectrum bias and limited external validation mean that broader independent validation is required before routine clinical use. Larger multicentre validation with standardised imaging protocols remains necessary to confirm generalisability.
Source: Insights into Imaging
Image Credit: iStock
References:
Xie R, Ai Y, Bao A et al. (2026) MRI- and report-based multimodal model with SHAP-based explanation for preoperative prediction of deep stromal invasion in early-stage cervical cancer. Insights Imaging; 17, 132.
