# Deep learning and dual-radiomics model incorporating brachytherapy applicator type to predict radiation-induced acute rectal injury in cervical cancer patients

## TL;DR

This study develops a model combining deep learning and radiomics to predict rectal injury caused by radiation in cervical cancer patients.

## Contribution

The novel integration of brachytherapy applicator type with deep learning and dual-radiomics improves prediction of radiation-induced rectal injury.

## Key findings

- The nomogram integrating multiple features achieved an external AUC of 0.803 for predicting rectal injury.
- Deep learning models outperformed dual-radiomics in external validation with an AUC of 0.773.
- Brachytherapy applicator type was significantly associated with acute rectal injury (P < 0.05).

## Abstract

•Predicting radiation-induced rectal injury is critical for cervical cancer prognosis.•Brachytherapy applicator type was associated with acute rectal injury (P < 0.05).•External area under curve was 0.773 for deep learning and 0.755 for dual-radiomics.•Nomogram achieved best area under curve of 0.810 (internal) and 0.803 (external).

Predicting radiation-induced rectal injury is critical for cervical cancer prognosis.

Brachytherapy applicator type was associated with acute rectal injury (P < 0.05).

External area under curve was 0.773 for deep learning and 0.755 for dual-radiomics.

Nomogram achieved best area under curve of 0.810 (internal) and 0.803 (external).

Radiation-induced acute rectal injury (RARI) is a common early toxicity after radiotherapy for cervical cancer (CC) and remains difficult to predict before treatment, which can adversely affect life quality of patients. We aimed to develop a combined dual-radiomics and deep learning (DL) model to improve the prediction of RARI in CC patients treated with radiotherapy.

This retrospective study included 200 CC patients from one hospital, randomly divided into training (n = 160), internal validation (n = 40) cohorts and external validation (n = 40) from another hospital. Patients were classified as RARI (CTCAE v5.0 grade ≥ 2) or Non-RARI (grade < 2). Radiomic and dosiomic features were extracted from CT images and dose distributions, and DL features were learned using 3D CNNs. The performance of radiomics, dosiomics, DL and hybrid features models for RARI prediction was compared using the receiver operating characteristic (ROC) curve with measurement of the area under the curve (AUC).

For radiomics combining dosiomics, XGBoost achieved the best performance with AUCs of 0.786 and 0.755 in internal and external validation cohorts, respectively. For DL, Resnet_with_CBAM achieved the best performance in the input of combining CT and dose distribution with AUCs of 0.786 and 0.773 in internal and external validation cohorts, respectively. Nomogram integrating radiomics, dosiomics, DL features, and clinical factor improved the AUC to 0.810, 0.803 in internal and external validation cohorts, respectively.

The nomogram integrating radiomics, dosiomics, DL, and clinical factors can improve the predictive performance for RARI in CC patients followed by radiotherapy.