Radiomics analysis for prediction of lymph node metastasis after neoadjuvant chemotherapy based on pretreatment MRI in patients with locally advanced cervical cancer

Jinjin Liu, Linxiao Dong, Xiaoxian Zhang, Qingxia Wu, Zihan Yang, Yuejie Zhang, Chunmiao Xu, Meiyun WangFrontiers in Oncology 2024 · DOI

TL;DR

Researchers have developed a multi-sequence MRI radiomics model that predicts whether cervical cancer has spread to lymph nodes following neoadjuvant chemotherapy (NACT). By combining textural features from pretreatment scans with clinical staging, the model achieves high predictive accuracy (AUC of 0.859 in validation), potentially allowing clinicians to skip unnecessary surgeries in favor of more effective chemoradiotherapy.

What Problem Does This Paper Address?

In locally advanced cervical cancer (LACC), doctors often use neoadjuvant chemotherapy (NACT) to shrink tumors before surgery. A critical factor in a patient's prognosis is lymph node metastasis (LNM). If the cancer remains in the lymph nodes after chemotherapy, radical surgery might be less effective than moving straight to chemoradiotherapy.

The challenge is that standard MRI is notoriously poor at identifying these "positive" nodes—it often misses cancer in normal-sized nodes (low sensitivity). Currently, there is no reliable way to use a patient's initial MRI to predict their lymph node status after they finish chemotherapy. This paper aims to fill that gap using radiomics—the extraction of thousands of quantitative features from medical images that are invisible to the human eye.

Methods

The study followed a classic radiomics pipeline but applied it to a complex multi-modal imaging set:

  1. Imaging Sequences: Features were extracted from three distinct MRI types: Sagittal T2-weighted (anatomy), Axial Diffusion-weighted (Ax-DWI, cellular density), and Sagittal T1-weighted contrast-enhanced (blood flow).
  2. Feature Engineering: Using the PyRadiomics package, the team extracted 2,264 features per patient, covering first-order statistics, shape, and complex textures (e.g., GLCM, GLSZM).
  3. Dimensionality Reduction: To avoid "overfitting" (finding patterns in noise), they used K-best and LASSO (Least Absolute Shrinkage and Selection Operator) methods to whittle thousands of features down to the most predictive few.
  4. Model Construction: They built a Random Forest classifier. They compared models based on single sequences against a "Combined Model" that fused radiomics with the clinical FIGO stage.
  5. Validation: The model was trained on data from one hospital and "stress-tested" on an independent validation set from a second hospital.

Datasets

The researchers conducted a retrospective study of 282 patients enrolled between 2013 and 2022:

  • Training Set (Center 1): 171 patients from Henan Province People’s Hospital.
  • External Validation Set (Center 2): 111 patients from Henan Cancer Hospital.
  • Ground Truth: Lymph node status was confirmed via pathology reports following radical hysterectomy (the "gold standard").

Key Insights and Ideas

  • DWI is the Strongest Signal: Among the single-sequence models, the Ax-DWI model performed best. This is likely because Diffusion-Weighted Imaging captures water fluidity and cell membrane integrity, which are highly sensitive markers for tumor aggressiveness and heterogeneity.
  • Texture Over Shape: Interestingly, the study found that tumor shape features were not significant predictors. Instead, textural features (which quantify how "busy" or "heterogeneous" an image is) were the primary drivers of the model. This suggests that the internal "chaos" of a tumor is more telling than its outward geometry.
  • Clinical-Radiomics Synergy: While radiomics is powerful, it didn't replace clinical intuition. Adding the FIGO stage (a standard clinical measure of cancer spread) to the radiomics features significantly boosted the model's performance in the validation set.
  • The "Net Benefit" Threshold: Decision Curve Analysis (DCA) showed that the combined model provides a higher "net benefit" than standard clinical staging alone, particularly when the clinical suspicion (threshold probability) of metastasis is below 80%.

Key Results

The Combined Model (MRI features + FIGO stage) outperformed all other approaches:

  • Training Set AUC: 0.889 (95% CI: 0.833–0.945)
  • Validation Set AUC: 0.859 (95% CI: 0.781–0.936)
  • Accuracy: 72.1% in the validation set, with a high specificity (96.0%), meaning the model is excellent at correctly identifying patients who do not have metastasis, potentially saving them from over-treatment.

By comparison, the FIGO stage alone only achieved an AUC of 0.697 in the training set, highlighting the massive upgrade provided by radiomic analysis.

Limitations and Future Work

  • Retrospective Nature: The data was collected after the fact. A prospective study (tracking patients in real-time) is needed to confirm these findings before clinical adoption.
  • The "Delta" Missing Link: This study used pretreatment scans to predict post-treatment outcomes. The authors note that "delta-radiomics"—analyzing how features change between a scan before NACT and a scan during NACT—might provide even more powerful predictive data.
  • Sample Size: While 282 patients is substantial for this niche, larger multi-center cohorts would help generalize the model across different MRI hardware and protocols.

Summary updated

Abstract

The pretreatment MRI-based radiomics model, integrating radiomics features from three sequences and clinical variables, exhibited superior performance in predicting LNM following NACT in patients with LACC.
BibTeX
@inproceedings{liu2024radiomics,
  author = {Jinjin Liu and Linxiao Dong and Xiaoxian Zhang and Qingxia Wu and Zihan Yang and Yuejie Zhang and Chunmiao Xu and Qingxia Wu and Meiyun Wang and Qingxia Wu and Meiyun Wang},
  title = {{Radiomics analysis for prediction of lymph node metastasis after neoadjuvant chemotherapy based on pretreatment MRI in patients with locally advanced cervical cancer}},
  booktitle = {Frontiers in Oncology},
  year = {2024},
  doi = {10.3389/fonc.2024.1376640},
}
Figure