Automated Prediction of Neoadjuvant Chemoradiotherapy Response in Locally Advanced Cervical Cancer Using Hybrid Model-Based MRI Radiomics

Hua Yang, Yinan Xu, Mohan Dong, Ying Zhang, Jie Gong, Dong Huang, Junhua He, Lichun Wei, Shigao Huang, Lina ZhaoDiagnostics 2023 · DOI

TL;DR

This paper introduces a novel hybrid machine learning model that combines MRI radiomics and clinical data to predict treatment response in locally advanced cervical cancer (LACC) patients undergoing neoadjuvant chemoradiotherapy (nCRT). The model achieves promising predictive performance, offering a potential tool for oncologists to personalize treatment strategies.

What Problem Does This Paper Address?

Cervical cancer (CC) remains a significant global health challenge, particularly in its locally advanced stages (LACC). While radiotherapy is a cornerstone treatment, patient responses vary significantly, with some experiencing residual disease or recurrence. Predicting a patient's sensitivity to neoadjuvant chemoradiotherapy (nCRT) before treatment is crucial for tailoring therapeutic strategies and improving outcomes. Current methods often rely on traditional imaging features or clinical parameters, which may not fully capture the complex biological and morphological characteristics of tumors that influence treatment response. This study aims to bridge this gap by developing an automated prediction model that leverages advanced imaging analysis (radiomics) and clinical data to forecast the pathological complete response (pCR) to nCRT.

Methods

The researchers developed a "hybrid model" that integrates three types of information:

  • Domain-Specific Radiomic Features: These are handcrafted features extracted from T2-weighted MRI images using the PyRadiomics toolkit. They capture quantitative characteristics of the tumor, such as shape, size, and texture.
  • Abstract Features from Deep Learning: A pre-trained VGG19 deep learning network, fine-tuned on the study's MRI data, was used to extract abstract, high-level features. This approach leverages the network's ability to learn complex patterns that might not be explicitly defined by traditional radiomics.
  • Clinical Parameters: Standard clinical information, including age, stage, pathological type, and lymph node status, was also incorporated.

These three feature sets were then combined and fed into an ensemble learning classifier. This ensemble approach uses a weighted voting strategy among multiple classifiers (Logistic Regression, Bayesian, KNN, and SVM) to aggregate their predictions, aiming for a more robust and accurate final prediction of pCR. The ensemble weights were adaptively determined based on the performance of each individual classifier on validation data.

Datasets

The study utilized a dataset of 138 locally advanced cervical cancer (LACC) patients treated at Xijing Hospital, Air Force Medical University, between January 2015 and December 2019. T2-weighted MRI images and corresponding clinical information were collected for each patient. The dataset was divided into a pathological complete response (pCR) group (74 patients) and a non-pCR group (64 patients). The paper does not mention the creation of a new benchmark dataset.

Key Insights and Ideas

  • The Power of Hybridization: The most significant insight is that combining domain-specific radiomic features, abstract deep learning features, and clinical parameters yields superior prediction performance compared to using any single feature type alone. This highlights the complementary nature of these data sources.
  • Deep Learning's Abstract Power: Abstract features extracted by the VGG19 network proved more effective for prediction than traditional domain-specific radiomic features, suggesting that deep learning can uncover subtle patterns indicative of treatment response.
  • Ensemble Learning for Robustness: Employing an ensemble classifier significantly improved prediction metrics, demonstrating that aggregating predictions from multiple models can lead to more stable and accurate outcomes, especially when dealing with complex medical data.
  • MRI's Predictive Potential: The study reinforces the value of T2-weighted MRI in capturing information relevant to predicting treatment response in cervical cancer, going beyond its traditional role in staging.

Key Results

The proposed hybrid model, integrating all three feature types and using ensemble learning, achieved an average AUC of 0.797 ± 0.023. This performance was notably better than models relying solely on clinical parameters (AUC 0.609), domain-specific features (AUC 0.674), or abstract features (AUC 0.760). The hybrid model also demonstrated improved accuracy (ACC), true positive rate (TPR), true negative rate (TNR), and precision. The ensemble classifier itself outperformed individual classifiers like Bayesian, Logistic Regression, KNN, and SVC when fed with the combined feature set.

Limitations and Future Work

The authors acknowledge two primary limitations:

  1. Dataset Size: The dataset comprised 138 patients, which is considered limited. While they employed rigorous cross-validation (five-fold repeated 20 times) to mitigate overfitting and ensure generalizability, they emphasize that expanding the sample size is crucial for further validation.
  2. Retrospective Nature and MRI Sequences: The study was retrospective, and the MRI protocols at the time did not include diffusion sequences. Therefore, only features from T2WI sequences were extracted. The authors suggest that incorporating features from other sequences, such as diffusion-weighted imaging, could potentially extract more representative features and improve model performance.

Future work could focus on larger, multi-center datasets and exploring the integration of additional MRI sequences and potentially other imaging modalities or genomic data to further enhance predictive accuracy and clinical utility.

Summary updated

Abstract

The proposed hybrid model can predict the radiotherapy response of patients with LACC, which might help radiation oncologists create personalized treatment plans for patients.
BibTeX
@inproceedings{yang2023automated,
  author = {Hua Yang and Yinan Xu and Mohan Dong and Ying Zhang and Jie Gong and Dong Huang and Junhua He and Lichun Wei and Shigao Huang and Lina Zhao},
  title = {{Automated Prediction of Neoadjuvant Chemoradiotherapy Response in Locally Advanced Cervical Cancer Using Hybrid Model-Based MRI Radiomics}},
  booktitle = {Diagnostics},
  year = {2023},
  doi = {10.3390/diagnostics14010005},
}
Figure