Application and Clinical Value of Machine Learning‐Based Cervical Cancer Diagnosis and Prediction Model in Adjuvant Chemotherapy for Cervical Cancer: A Single‐Center, Controlled, Non‐Arbitrary Size Case‐Control Study

Yan Wang, Lidan Shen, Jun Jin, Guohua WangContrast 2022 · DOI

TL;DR

This study developed and validated a machine learning model using MRI radiomics to differentiate early-stage cervical cancer (Stage IA CC) from precancerous lesions (HSIL), achieving high diagnostic accuracy. The model also showed promise in monitoring chemotherapy response, offering a non-invasive tool to guide treatment decisions and potentially reduce the need for invasive procedures.

What Problem Does This Paper Address?

Accurate and early diagnosis of cervical cancer (CC) is crucial for effective treatment and improved patient outcomes, especially in regions with limited medical resources. Differentiating between Stage IA CC and high-grade squamous intraepithelial lesions (HSIL) can be challenging, even with conventional MRI, as the visual differences can be subtle. Current diagnostic methods often rely on invasive procedures. This paper aims to address this by leveraging machine learning and radiomics from MRI data to create a more precise and less invasive diagnostic tool for these early-stage lesions and to assess its value in predicting response to neoadjuvant chemotherapy (NACT).

Methods

The researchers employed a case-control study design, retrospectively analyzing MRI data from 46 patients with Stage IA CC and 55 patients with HSIL. The core of their approach involved:

  • Radiomics Feature Extraction: MRI images (OSag-T2WI, OAx-T1WI, OAx-T2FS) were preprocessed, and a region of interest (ROI) encompassing the cervix was delineated. From these ROIs, a comprehensive set of 1409 radiomic features were extracted, including first-order, shape, and texture features.
  • Feature Selection: To reduce dimensionality and identify the most informative features, three methods were used: Variance Threshold, SelectKBest, and LASSO regression.
  • Machine Learning Model: An arbitrary forest (RF) model was chosen for classification. Four different RF models were built, each utilizing features from different MRI sequences or combinations thereof.
  • Model Evaluation: The models were trained and tested using a 4:1 ratio split. Diagnostic performance was assessed using Receiver Operating Characteristic (ROC) curves, with key metrics including Area Under the Curve (AUC), sensitivity, and specificity.
  • Chemotherapy Response Monitoring: The study also analyzed changes in ADC (Apparent Diffusion Coefficient) values before and after NACT in patients with CC to evaluate the model's potential for monitoring treatment efficacy.

Datasets

The study utilized a dataset of 46 patients with Stage IA CC and 55 patients with HSIL, collected retrospectively from August 2019 to August 2021 at a single center. The data was split into training (IA=37, HSIL=44) and testing (IA=9, HSIL=11) sets. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Subtle Differences, Powerful Data: The study highlights that even subtle differences in MRI images, imperceptible to the human eye, can be captured and leveraged by radiomics and machine learning to distinguish between precancerous and early cancerous lesions.
  • Multi-Sequence Fusion is Key: Combining radiomic features from multiple MRI sequences (specifically OSag-T2WI and OAx-T2FS) yielded the best diagnostic performance, suggesting that a comprehensive view of tissue characteristics across different imaging modalities is more powerful than any single sequence alone.
  • ADC as a Dynamic Biomarker: Changes in ADC values before and after chemotherapy appear to correlate with treatment response, offering a potential non-invasive method to dynamically monitor treatment efficacy and adjust strategies.
  • Reducing Invasive Procedures: The development of such a predictive model could significantly reduce the need for invasive diagnostic procedures, improving patient comfort and potentially lowering healthcare costs.

Key Results

  • ADC Value Differentiation: The ADC value of CC tissues was significantly lower than that of normal cervical tissues (P < 0.05). The ROC analysis for differentiating CC from normal cervix showed an AUC of 0.838, with a sensitivity of 92.21% and specificity of 9.48% at an optimal cutoff value.
  • Chemotherapy Response: In patients who responded to NACT, ADC values were significantly higher after the second chemotherapy compared to before treatment and after the first chemotherapy (P < 0.05). Ineffective patients showed no significant changes in ADC values.
  • Best Performing Model: The combined OSag-T2WI & OAx-T2FS arbitrary forest model achieved the highest diagnostic efficiency. For the test set, this model reported an AUC of 0.894 (95% CI: 0.75-1.00) and an accuracy rate of 0.835.
  • OAx-T1WI Underperformance: The model based solely on OAx-T1WI features performed poorly, with an AUC of 0.513 in the test set, indicating it had no practical clinical significance. This was attributed to the dominance of first-order features, which are less robust than texture features.

Limitations and Future Work

  • Single-Center Study: The study was conducted at a single center, which may limit the generalizability of the findings to diverse patient populations and clinical settings.
  • Small Sample Size: While the study included a reasonable number of patients, the authors acknowledge that the sample size, particularly for the test set and for analyzing different stages of CC, is relatively small. Further studies with larger, multi-center datasets are needed.
  • ADC Threshold for Staging: The study did not investigate the specific ADC threshold values for diagnosing different stages of CC, suggesting this as an area for future research.
  • Dynamic Monitoring Potential: The authors suggest that DWI can dynamically monitor chemotherapy efficacy, indicating potential clinical application value, but further validation is required.
  • Feature Selection Impact: The authors speculate that the choice of ROI delineation might have influenced the performance of certain texture features like GLDM, suggesting further investigation into optimal segmentation strategies.

Summary updated

Abstract

A CC diagnosis and prediction model based on machine learning can better distinguish stage IA CC from HSIL in the absence of clear lesions, which is of great significance for reducing invasive examination before surgery, guiding surgical procedures and adjuvant chemotherapy for CC.
BibTeX
@inproceedings{wang2022application,
  author = {Yan Wang and Lidan Shen and Jun Jin and Guohua Wang},
  title = {{Application and Clinical Value of Machine Learning‐Based Cervical Cancer Diagnosis and Prediction Model in Adjuvant Chemotherapy for Cervical Cancer: A Single‐Center, Controlled, Non‐Arbitrary Size Case‐Control Study}},
  booktitle = {Contrast},
  year = {2022},
  doi = {10.1155/2022/2432291},
}
Figure