Moderate evidence
More
CT-based model told benign from malignant bowel lesions with 0.957 AUC in a two-centre study
Study title: Multicenter evaluation of radiomics and multimodal artificial intelligence for colorectal cancer diagnosis and survival prediction
- Evidence type
- Controlled or validated comparison
- Sample
- n = 879
- Sources
- 1 sourcePeer-reviewed
- Published
- 1 min read
- Truvace record
- TRV-2026-3065cite or check it
Study title: Multicenter evaluation of radiomics and multimodal artificial intelligence for colorectal cancer diagnosis and survival prediction
Researchers retrospectively analysed 879 participants from two centres, 741 with locally advanced colorectal cancer and 138 with benign colorectal lesions. Models were trained on pretreatment contrast-enhanced CT scans and clinical data. A clinical-radiomics nomogram reached a test AUC of 0.957 for telling benign from malignant lesions.
Read the study abstract
Abstract Reliable preoperative classification and outcome prediction could improve management of locally advanced colorectal cancer. We retrospectively studied 879 participants from two centers, including 741 patients with locally advanced colorectal cancer and 138 with benign colorectal lesions. Clinical, radiomics, deep learning, and early fusion models were trained at center A ( n = 615) and independently tested at center B ( n = 264) using pretreatment contrast-enhanced computed tomography and clinical data.…

The gain
In a retrospective study of 879 patients at two centres, a clinical-radiomics model trained at one centre separated malignant from benign colorectal lesions with a test AUC of 0.957 at the second centre.
In brief
For overall survival, radiomics reached test AUCs of 0.874 at three years and 0.898 at five years. A deep learning fusion model scored 0.870 and 0.899, essentially matching it.
The authors conclude that CT radiomics is a promising digital biomarker for stratifying colorectal cancer, while early deep learning fusion adds limited value over an optimised radiomics baseline.
Main points
- The study included 879 participants from two centres: 741 with locally advanced colorectal cancer and 138 with benign lesions.
- Clinical, radiomics, deep learning and early fusion models were trained at center A (n = 615).
- Models were independently tested at center B (n = 264).
- Inputs were pretreatment contrast-enhanced CT and clinical data.
- The clinical-radiomics nomogram reached a test AUC of 0.957 for benign-malignant classification.
- Three-year survival test AUC was 0.874 for radiomics and 0.870 for fusion; five-year was 0.898 and 0.899.
The full story
The study was retrospective and covered two centres. Center A supplied 615 participants for training, and center B supplied 264 for independent testing, which gives the results an external check.
The models compared were clinical, radiomics, deep learning and early fusion, all built on pretreatment contrast-enhanced CT and clinical data. The aim was better preoperative classification and outcome prediction in locally advanced colorectal cancer.
The clinical-radiomics nomogram scored a test AUC of 0.957 for benign versus malignant. Survival predictions were similar across approaches: radiomics 0.874 and 0.898, fusion 0.870 and 0.899 at three and five years.
Sources
1 on file- npj Digital Medicinedoi:10.1038/s41746-026-03434-3
Talk it over
The clinical-radiomics nomogram reached a test AUC of 0.957 at center B after training at center A, in a retrospective study. How much weight would you put on that result, and why?
CT-based model told benign from malignant bowel lesions with 0.957 AUC in a two-centre studyhttps://truvace.com/item/multicenter-evaluation-of-radiomics-and-multimodal-artificial-intelligence-for-c
The clinical-radiomics nomogram reached a test AUC of 0.957 at center B after training at center A, in a retrospective study. How much weight would you put on that result, and why?
Answer in the threadThe study included 741 patients with locally advanced colorectal cancer and 138 with benign lesions. Which kinds of patients or scanners could be missing from that mix, and how might that change the result?
Answer in the threadThe deep learning fusion model scored 0.870 and 0.899 for survival, close to radiomics at 0.874 and 0.898. Which model would you pick for a hospital to use, and what would drive your choice?
Answer in the thread
If you have more time
What would a prospective study at more centres need to show before doctors used the nomogram to sort benign from malignant lesions before surgery?
Answer in the threadIf you were a radiologist at center B checking the nomogram's results, what would you compare them against, and who else would you want involved in that check?
Answer in the thread
ace
What readers say