An Automated Radiomics Framework for Postoperative Survival Prediction in Colorectal Liver Metastases using Preoperative MRI
New AI analyzes pre-surgery MRI to predict patient survival, potentially preventing unnecessary high-risk operations.
A multi-institutional research team has published a new AI framework that automates the prediction of survival outcomes for patients with colorectal cancer that has spread to the liver (CRLM). The system analyzes standard pre-operative MRI scans to forecast how long a patient might live after curative-intent liver surgery. This addresses a critical clinical challenge, as patient outcomes after these major operations are highly variable, and surgeons currently lack reliable tools to identify which patients will truly benefit from the procedure.
The framework, detailed in a preprint on arXiv, consists of two core AI pipelines. First, a segmentation pipeline called SAMONAI extends Meta's Segment Anything Model (SAM) into 3D to automatically identify and outline the liver, spleen, and individual liver tumors from MRI data. This step achieved a high Dice score of 0.78 for tumor segmentation. These precise segmentations then feed into a second, novel radiomics pipeline named SurvAMINN. This multiple instance neural network extracts quantitative features from the tumors and learns to predict time-to-event survival data from historical patient records.
In testing on data from 227 patients, the combined system predicted survival with a Concordance Index (C-index) of 0.69, outperforming established clinical biomarkers. The model is designed to emphasize identifying high-risk metastases that are likely to lead to poor outcomes. By providing an automated, objective assessment from routine scans, this tool could help clinicians personalize therapy plans and avoid major surgeries for patients unlikely to see a survival benefit.
- The 'SAMONAI' framework extends Meta's Segment Anything Model for 3D, anatomy-aware segmentation of liver tumors from MRI, achieving a Dice score of 0.78.
- Its survival prediction component, 'SurvAMINN', is a custom neural network that achieved a C-index of 0.69 on a retrospective cohort of 227 colorectal liver metastasis patients.
- The fully automated system analyzes pre-operative scans to predict post-surgery survival, aiming to prevent non-beneficial, high-risk hepatectomy procedures.
Why It Matters
This AI could transform surgical planning for metastatic cancer, providing data-driven guidance to avoid major, invasive surgeries for patients who won't benefit.