Predicting local persistence/recurrence after radiation therapy for head and neck cancer from PET/CT using a multi-objective, multi-classifier radiomics model

Qiongwen Zhang, Kai Wang, Zhiguo Zhou, Genggeng Qin, Lei Wang, Ping Li, David J. Sher, Steve Jiang, Jing WangFrontiers in Oncology 2022 · DOI

TL;DR

This paper introduces a novel multi-objective, multi-classifier radiomics model that leverages post-treatment PET/CT scans and clinical data to predict local persistence or recurrence of head and neck squamous cell carcinoma (HNSCC) after radiation therapy. The model achieves high accuracy, offering a promising tool for personalized patient management and early detection of cancer recurrence.

What Problem Does This Paper Address?

Head and neck squamous cell carcinoma (HNSCC) is a significant cancer with a substantial risk of local persistence or recurrence (P/R) even after curative intent radiation therapy, affecting 25-50% of patients. Early and accurate prediction of P/R is crucial for tailoring patient management, allowing for intensified treatment or more frequent surveillance for high-risk individuals, while potentially sparing low-risk patients from unnecessary procedures. Current surveillance methods, like FDG-PET/CT, face challenges due to treatment-related inflammation, which can lead to false-positive interpretations, especially within the first 12 weeks post-treatment. This paper aims to address this by developing a radiomics-based model that can predict P/R using post-treatment PET/CT scans, potentially distinguishing between true recurrence and inflammatory changes.

Methods

The researchers developed a sophisticated multi-objective, multi-classifier, and multi-modality radiomics model. Key methodological aspects include:

  • Radiomic Feature Extraction: 257 radiomic features were extracted from post-treatment PET/CT scans within the gross tumor volume (GTV). These features capture various aspects of tumor heterogeneity, including geometry, intensity, and texture (GLCMs).
  • Multi-Objective Optimization: The model was trained using a multi-objective approach, simultaneously optimizing for sensitivity, specificity, and feature sparsity. Feature sparsity was introduced as an additional objective to encourage the selection of a more concise and potentially more robust set of features. An immune algorithm (IMIA) was employed for this optimization.
  • Multi-Classifier Strategy: To enhance prediction robustness, the model integrated multiple classifiers (logistic regression, discriminant analysis, and support vector machine) for each modality. These individual classifier outputs were then fused.
  • Multi-Modality Fusion: The model combined information from PET radiomics, CT radiomics, and clinical data. The output probabilities from models trained on each modality were fused using an evidential reasoning (ER) method to generate a final prediction.
  • Feature Pre-selection: Before model training, a minimal-redundancy-maximal-relevance (mRMR) method was used to pre-select 50 features for each imaging modality.

Datasets

The study utilized a retrospective dataset of 328 patients with HNSCC treated with definitive radiation therapy at a single institution. Of these, 69 patients experienced local P/R, and 259 did not. The dataset was divided into a training cohort (262 patients) and a validation cohort (66 patients) using a 4:1 case-control matching fashion. No new benchmarks were introduced; the focus was on validating the proposed model on this existing patient data.

Key Insights and Ideas

  • Post-Treatment Radiomics for P/R Prediction: The study demonstrates that radiomic features extracted from post-treatment PET/CT scans can effectively predict local persistence or recurrence, moving beyond the common practice of using pre-treatment data for prognostication.
  • Value of Multi-Objective Optimization with Sparsity: Incorporating feature sparsity as an objective significantly reduced the number of selected radiomic features while improving model performance, suggesting that parsimonious models can be more effective.
  • Synergy of Multi-Modality Fusion: Combining PET, CT, and clinical data yielded the best predictive performance, highlighting the complementary nature of these information sources for HNSCC recurrence prediction.
  • Potential to Differentiate from Inflammation: The model's ability to predict recurrence even at relatively early post-treatment stages (median 114 days) suggests it might be able to distinguish true tumor recurrence from post-radiation inflammation, a key challenge with standard PET/CT interpretation.

Key Results

The developed multi-objective, multi-classifier, multi-modality radiomics model (CT+PET+Clinic) achieved the following performance on the validation cohort:

  • Sensitivity: 93%
  • Specificity: 83%
  • Accuracy: 85%
  • AUC: 0.94

This performance was superior to single-modality models and models that did not incorporate feature sparsity. The CT+PET+Clinic model also demonstrated a significant ability to differentiate between high- and low-risk patient groups for local P/R free survival, as shown by Kaplan-Meier curves.

Limitations and Future Work

The authors acknowledge several limitations:

  • Single Institution Data: The model was trained and validated on data from a single institution, which may limit its generalizability due to variations in treatment planning, follow-up, and assessment protocols.
  • Additive Benefit of Clinical Features: While the CT+PET+Clinic model trended towards better performance, the study could not definitively prove an additive benefit of clinical features over PET/CT radiomics alone, possibly due to the limited patient numbers.
  • Post-Treatment PET/CT Only: The model exclusively used post-treatment PET/CT images. Future work could explore incorporating other imaging modalities (e.g., MRI, CBCT) and data collected at different time points (e.g., pre-treatment PET/CT, intra-treatment imaging).

Future work should focus on:

  • Multicenter Validation: Conducting a large-scale, multicenter analysis to validate the model's clinical applicability and robustness.
  • External Validation: Testing the model on independent external datasets.
  • Integrating More Data: Exploring the inclusion of dosimetric factors, pre-treatment PET/CT images, and other imaging modalities like PET/MRI for a more comprehensive prediction model.

Summary updated

Abstract

Our study demonstrates the feasibility of employing a multi-objective, multi-classifier radiomics model with PET/CT radiomic features and clinical data to predict outcomes for patients with HNSCC after radiation therapy. The proposed prediction model shows the potential to detect cancer local P/R early after radiation therapy.
BibTeX
@inproceedings{zhang2022predicting,
  author = {Qiongwen Zhang and Kai Wang and Zhiguo Zhou and Genggeng Qin and Lei Wang and Ping Li and David J. Sher and Steve Jiang and Jing Wang},
  title = {{Predicting local persistence/recurrence after radiation therapy for head and neck cancer from PET/CT using a multi-objective, multi-classifier radiomics model}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.955712},
}
Figure