KEY POINTS
- This retrospective single-center study included 650 patients with cervical cancer treated between 2019 and 2023 with postoperative radiotherapy or cisplatin-based chemoradiotherapy. Treatment delivered 50.0–50.4 Gy in 1.8–2.0 Gy fractions, with concurrent cisplatin 40 mg/m² weekly where indicated.
- Imaging deliberately included two different platforms. 545 patients had simulation CT on a Siemens scanner, which supplied the training and internal-test datasets, while 105 patients underwent fan-beam CT on a United Imaging uRT-linac and served as an independent device-level test cohort.
- Clinically significant radiation proctitis was defined as CTCAE grade ≥2. Its incidence was 33.3% in training, 27.5% in the internal test set and 27.6% in the independent-device test set, providing relatively similar event rates across datasets.
- The C2RISNet segmentation architecture combined convolutional layers, attention modules and Mamba state-space modeling to segment CTV, bladder and rectum automatically. Mean Dice similarity coefficient reached 90.76% internally and 85.33% on the independent device, outperforming the tested U-Net, Attention U-Net, Swin U-Net, TransUNet and Swin-UMamba alternatives.
- For toxicity prediction, the investigators combined conventional radiomics, features extracted by a self-supervised vision foundation model and clinical/dosimetric data. The final multimodal model achieved AUC 0.882, accuracy 89.0%, sensitivity 93.3% and specificity 87.3% on the internal test set.
- Performance remained relatively stable when transferred to the second imaging device: AUC 0.865, accuracy 85.7%, sensitivity 79.3% and specificity 88.2%. By comparison, the clinical-data-only model achieved AUCs of only 0.710 and 0.716 in the internal and independent-device sets, respectively.
- SHAP analysis identified rectal dosimetric variables alongside foundation-model and radiomic features as major contributors to predictions. Calibration was also favorable, with Brier scores of 0.099 internally and 0.139 on the independent device, but both devices came from the same institution and the model has not been prospectively tested to guide treatment-plan modification.
CLINICAL TAKEAWAY
The interesting part is not simply another radiomics AUC: the same pipeline performed automatic pelvic segmentation and retained toxicity-prediction performance on a different CT platform. Still, this remains retrospective single-center development and validation; whether using the model to alter rectal dose actually reduces proctitis requires prospective external testing.