Moderate evidence
More
Scan-based AI beat radiologists' measures at spotting chemo responders in a 96-patient Ewing sarcoma study
Study title: Can radiomics outperform radiologists in predicting Ewing sarcoma response to neoadjuvant chemotherapy from longitudinal MRI?
- Evidence type
- Observational study
- Sample
- n = 96
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2893cite or check it
Study title: Can radiomics outperform radiologists in predicting Ewing sarcoma response to neoadjuvant chemotherapy from longitudinal MRI?
A retrospective multicenter study of 96 people with Ewing sarcoma, a bone and soft-tissue cancer, tested whether machine learning on MRI scans can predict how well tumours respond to chemotherapy. The best radiomics model, built on changes in T1 fat-saturated contrast-enhanced images, reached an AUROC of 0.62, against 0.58 for the best model based on radiologist readings.
Read the study abstract
This retrospective, multicenter study evaluated whether MRI-based radiomics can predict the histological response to neoadjuvant chemotherapy in Ewing sarcoma patients and compared their performance to that of radiological findings. Definitive histological assessment is only possible during surgery, potentially delaying therapy changes and increasing metastatic risk; therefore, a positive outcome would enable early adaptations. Radiomics features were extracted from pre- and post-neoadjuvant (after first chemoth…

The gain
In a retrospective study of 96 Ewing sarcoma patients, MRI radiomics models identified chemotherapy responders with an AUROC of 0.62, above 0.58 for the best model built on radiologist readings and 0.53 for a clinical baseline.
In brief
Definitive response can only be judged from tissue after surgery, which can delay changes to treatment and may raise metastatic risk. An earlier prediction from scans could allow therapy to be adapted sooner. The authors describe the identification of responders as moderate and the findings as exploratory, complementary to standard size-based assessment, and call for larger prospective trials.
Main points
- The study included 96 patients (mean age 19.9 ± 11.7 years; 41 women) treated for Ewing sarcoma between 2004 and 2020.
- Radiomics features were taken from scans before chemotherapy and after the first chemotherapy cycle.
- Four machine learning models were trained with nested cross-validation to predict histological response.
- The best radiomics model scored an AUROC of 0.62 (95% CI 0.59 to 0.65), using low gray level intensity changes in T1 fat-saturated contrast-enhanced sequences.
- The best radiologist-feature model, based on change in maximum diameter of extratumoral soft tissue components, scored 0.58 (95% CI 0.56 to 0.60).
- Radiomics models using only pre-treatment T2-weighted data scored 0.61, against 0.53 for the clinical baseline.
The full story
Researchers extracted radiomics features from MRI scans taken before neoadjuvant chemotherapy and after the first cycle in patients treated from 2004 to 2020 at multiple centers. They compared four machine learning models with models trained on semantic features from expert radiological readings, measuring performance by AUROC.
The top radiomics model used low gray level intensity changes in contrast-enhanced T1 fat-saturated scans and scored 0.62. The best radiological model relied on the relative change in maximum diameter of extratumoral soft tissue components and scored 0.58. Pre-treatment T2-weighted radiomics scored 0.61, while the clinical baseline scored 0.53.
The authors say radiomics may capture non-size tissue signals that could guide personalized treatment adaptations and reduce delays for non-responders. They describe the findings as exploratory and say larger studies within a multimodal framework are needed to validate them.
Sources
1 on file- BMC Cancerdoi:10.1186/s12885-026-17332-0
Talk it over
The best radiomics model reached an AUROC of 0.62, against 0.58 for the best model built on radiologist readings. How would you weigh that gap, given the authors call the identification of responders moderate?
Scan-based AI beat radiologists' measures at spotting chemo responders in a 96-patient Ewing sarcoma studyhttps://truvace.com/item/can-radiomics-outperform-radiologists-in-predicting-ewing-sarcoma-response-to-ne
The best radiomics model reached an AUROC of 0.62, against 0.58 for the best model built on radiologist readings. How would you weigh that gap, given the authors call the identification of responders moderate?
Answer in the threadThe 96 patients were treated between 2004 and 2020 and the analysis was retrospective. How might that span of years and scanning at multiple centers affect what the models learned?
Answer in the threadWhat result in a larger prospective trial would make you willing to change a young patient's chemotherapy plan based on a scan-based prediction after the first cycle?
Answer in the thread
If you have more time
Definitive response can only be judged from tissue after surgery. Who should decide how much earlier information from scans is worth, given the risk of delayed treatment changes?
Answer in the threadPre-treatment T2-weighted radiomics scored 0.61, against 0.53 for the clinical baseline. What would you ask the researchers about why those two results differ so much?
Answer in the thread
ace
What readers say