Machine Learning-Based Multiparametric Magnetic Resonance Imaging Radiomics Model for Preoperative Predicting the Deep Stromal Invasion in Patients with Early Cervical Cancer

Haowen Yan, Gaoting Huang, Zhihe Yang, Yirong Chen, Zhiming XiangJ. Imaging Inform. Medicine 2024 · DOI

TL;DR

This study introduces a machine learning model that combines MRI radiomics and clinical data to accurately predict deep stromal invasion in early cervical cancer patients. The developed model achieved a high AUC of 0.969, outperforming models based on radiomics or clinical data alone, offering a promising non-invasive tool for improved diagnosis and timely treatment.

What Problem Does This Paper Address?

Deep stromal invasion is a critical factor influencing treatment decisions and prognosis for early cervical cancer patients. Accurately identifying this invasion before surgery (radical hysterectomy) is crucial for personalized treatment planning. While MRI is a standard diagnostic tool, its visual interpretation can be subjective and prone to inter-observer variability. Radiomics, which extracts quantitative features from medical images, offers a more objective approach. However, previous radiomics-based models for deep stromal invasion prediction have shown room for improvement. This paper aims to bridge this gap by developing and evaluating a machine learning model that integrates multiparametric MRI radiomics with clinical data to achieve a more accurate and reliable preoperative prediction of deep stromal invasion.

Methods

The researchers employed a machine learning approach using the LightGBM algorithm to construct three predictive models:

  1. Radiomics Model: Utilized radiomics features extracted from MRI scans.
  2. Clinical Model: Used only clinical variables.
  3. Combined Model: Integrated both radiomics features and clinical predictors.

Key steps in their methodology included:

  • Radiomics Feature Extraction: A semi-automatic method using a 3D region growing algorithm (GrowCut) was used to segment the region of interest (ROI) on MRI images. A total of 2632 radiomics features were extracted using the "PyRadiomics" package.
  • Feature Selection: A multi-stage process was used:
  • Statistical significance (P<0.05) was used to filter features.
  • Pearson correlation coefficient was applied to remove highly correlated features.
  • Analysis of Variance (ANOVA) selected the top 15 features with high variance.
  • Least Absolute Shrinkage and Selection Operator (LASSO) regression, combined with five-fold cross-validation, was used for final feature selection.
  • Clinical Predictor Identification: Univariable and multivariable logistic regression analyses were performed to identify significant clinical predictors.
  • Model Training and Validation: The dataset was split into training (70%) and testing (30%) sets. Three LightGBM models were trained on the training set and then validated on the testing set. Hyperparameter optimization was performed using Optuna.
  • Performance Evaluation: Models were evaluated using metrics such as Area Under the Curve (AUC), F1 score, accuracy, sensitivity, specificity, and negative predictive value (NPV). Receiver Operating Characteristic (ROC) curves and Kolmogorov-Smirnov (KS) curves were also generated.

Datasets

The study utilized data from 229 patients with early cervical cancer who underwent radical hysterectomy (RH) combined with pelvic lymph node dissection (PLND). These patients were recruited from The First Affiliated Hospital of Jinan University. The dataset was split into a training set of 160 patients and a testing set of 69 patients. No new benchmark datasets were introduced; the study used existing patient data.

Key Insights and Ideas

  • Radiomics Power: The study strongly supports the utility of radiomics in capturing subtle imaging characteristics that are predictive of deep stromal invasion, outperforming clinical data alone.
  • Synergistic Combination: The most significant insight is that combining radiomics features with clinical predictors yields the best predictive performance. This suggests that neither data source alone is sufficient, but their integration unlocks superior diagnostic power.
  • LightGBM Efficacy: The use of LightGBM, a gradient boosting framework, proved effective in building accurate predictive models, demonstrating its suitability for complex medical image analysis tasks.
  • Non-invasive Prediction: The development of a non-invasive prediction model is a major step forward, potentially reducing the need for invasive procedures or improving the accuracy of pre-operative assessments.
  • Clinical Relevance: The identified predictors (e.g., FIGO staging, SCC-Ag, and specific radiomics features) offer actionable insights for clinicians, guiding their diagnostic and therapeutic strategies.

Key Results

  • Combined Model Superiority: The combined radiomics and clinical predictor model achieved the highest AUC in the training set (0.969, 95% CI 0.947–0.990) and the testing set (0.914, 95% CI 0.848–0.980).
  • Radiomics Model Performance: The radiomics-only model also showed strong performance, with an AUC of 0.951 (95% CI 0.922–0.980) in the training set and 0.882 (95% CI 0.806–0.959) in the testing set.
  • Clinical Model Performance: The clinical predictor model had a lower AUC in both training (0.769, 95% CI 0.703–0.835) and testing (0.767, 95% CI 0.663–0.870) sets, highlighting the added value of radiomics.
  • KS Test Results: The KS test values indicated strong risk differentiation ability for all models, with the combined model showing the highest KS value (0.69).
  • Significant Predictors: Clinical predictors like FIGO staging (IIA, IB, IIB) and SCC-Ag were found to be significantly associated with deep stromal invasion.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single-Center Study: The data was collected from a single institution, which may introduce selection bias and limit the generalizability of the findings.
  • MRI Image Variability: MRI images were acquired from different devices, potentially impacting radiomics feature stability. While normalization was performed, this remains a consideration.

Future work suggested includes:

  • Deep Learning Integration: Exploring more advanced deep learning methods, such as automated in-depth feature learning algorithms and deep convolutional neural networks, to potentially construct more reliable prediction models.
  • Multi-center Validation: Conducting well-designed, multi-center studies to validate the findings and enhance the generalizability of the developed models.
  • Broader Application: Investigating the application of similar machine learning and radiomics approaches for other aspects of cervical cancer diagnosis, prognosis, and treatment selection.

Summary updated

BibTeX
@article{yan2024machine,
  author = {Haowen Yan and Gaoting Huang and Zhihe Yang and Yirong Chen and Zhiming Xiang},
  title = {{Machine Learning-Based Multiparametric Magnetic Resonance Imaging Radiomics Model for Preoperative Predicting the Deep Stromal Invasion in Patients with Early Cervical Cancer}},
  journal = {J. Imaging Inform. Medicine},
  year = {2024},
  doi = {10.1007/s10278-023-00906-w},
}
Figure