Radiomics applied to Ultrasound for predicting treatment response in patients with Locally Advanced Cervical Cancer: a retrospective multicenter (RU-LACC) study

Summarise with VIRAT
Authors: VALENTINA DOLCI
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Objective: The aim of the present study is to develop and validate ultrasound-based machine learning models including radiomic features to predict treatment response in patients with locally advanced cervical cancer (LACC). Methods: This retrospective study included LACC patients treated with either neoadjuvant chemoradiotherapy followed by surgery (N-CTRT) or exclusive concurrent chemoradiotherapy (E-CTRT). Treatment response was assessed histologically in N-CTRT and by MRI or PET/CT at 3–6 months in E-CTRT. Responders had microscopic or no residual disease after N-CTRT or complete response on imaging after E-CTRT; all others were non-responders. Pre-treatment ultrasound images were manually contoured, and radiomic features were extracted from the tumor regions. The dataset was split into training (75%) and validation (25%) sets. Feature selection and model training were performed using various algorithms. Two models were developed: a radiomic-only model and a radiomic-clinical-ultrasound model integrating radiomic, clinical, and ultrasound variables. Performance was evaluated by the area under the receiver operating characteristic curve (AUC-ROC). Results: Among 222 patients, 157 were responders and 65 non-responders. Of 74 extracted radiomic features, Recursive Feature Elimination (RFE) identified two key variables (F_szm.sze and F_szm.zsnu.norm) used to build the radiomic-only model. For clinical and ultrasound variables, the mRMR (Maximum Relevance Minimum Redundancy) algorithm identified lymph node involvement (detected by ultrasound, MRI, or PET/CT) and maximum tumor diameter measured on ultrasound as the most relevant predictors. Both models were trained using the CatBoost classifier. The radiomic-only model achieved an AUC of 0.80 (95% CI 0.72–0.86) in the training set and 0.69 (95% CI 0.54–0.85) in the validation set, with corresponding sensitivities of 0.86 (95% CI 0.76-0.96) and 0.75 (95% IC 0.54-0.96) and specificity of 0.59 (95% CI 0.50-0.68) and 0.53 (95% CI 0.37-0.68) at its best cutoff (23% malignancy risk based on Youden’s index). The radiomic-clinical-ultrasound model achieved an AUC of 0.87 (95% CI, 0.81–0.92) in the training set and 0.81 (95% CI, 0.67–0.92) in the validation set, with corresponding sensitivities of 0.80 (95% CI, 0.68–0.91) and 0.75 (95% CI, 0.54–0.96), and specificities of 0.79 (95% CI, 0.71–0.86) and 0.68 (95% CI, 0.53–0.82) at the optimal cutoff (28% malignancy risk based on Youden’s index). Conclusion: Although our models demonstrated moderate performance, we believe this remains a promising area of research, as no highly accurate tool currently exists for assessing tumor response in patients with LACC. Larger, multicenter, and prospective studies on patients with homogenous treatment strategy and outcome are essential to further explore the potential of radiomics in this field.

Fields of Study & Subjects

Radiomics and Machine Learning in Medical Imaging Endometrial and Cervical Cancer Treatments MRI in cancer diagnosis

Publication details

Journal / source
Electronic Theses and Dissertations Repository (University of Pisa)
Publication date
22 January 2026
Language
IT
Document type
Thesis