Background and objective

In this study, we have analysed pretreatment positron-emission tomography/ computed tomography (PET/CT) images of head and neck squamous cell carcinoma (HNSCC) patients. We have used a publicly available dataset for our analysis. The clinical features of the patient, PET quantitative parameters, and textural indices from pretreatment PET-CT images are selected for the study. The main objective of the study is to use classifiers to predict the outcome for HNSCC patients and compare the performance of the model with the conventional statistical model (CoxPH).

Methods

We have applied a 40% fixed SUV threshold method for tumour delineation. Clinical features of each patient are provided in the dataset, and other features are calculated using LIFEx software. For predicting the outcome, we have implemented three classifiers - Random Forest classifier, Gradient Boosted Decision tree (GBDT) and Decision tree classifier. We have trained each model using 93 data points and test the model performance using 39 data points. The best model - GBDT is chosen based on the performance metrics.

Results

It is observed that typically three features: MTV (Metabolic tumour Volume), primary tumour site and GLCM_correlation are significant for prediction of survival outcome. For testing cohort, GBDT achieves a balanced accuracy of 88%, where conventional statistical model reported a balanced accuracy of 81.5%.

Conclusions

The proposed classifier achieves higher accuracy than the state of the art technique. Using this classifier we can estimate the HNSCC patient’s outcome, and depending upon the outcome treatment policy can be selected.

中文翻译：

---

基于HNSCC患者的临床特征和18FDG-PET / CT预处理对生存结果的预测。

背景和目标

在这项研究中，我们分析了头颈部鳞状细胞癌（HNSCC）患者的治疗前正电子发射断层扫描/计算机断层扫描（PET / CT）图像。我们已使用可公开获取的数据集进行分析。选择患者的临床特征，PET定量参数和来自预处理PET-CT图像的质地指数进行研究。该研究的主要目的是使用分类器预测HNSCC患者的预后，并将该模型的性能与常规统计模型（CoxPH）进行比较。

方法

我们已将40％的SUV阈值固定方法用于肿瘤定位。数据集中提供了每个患者的临床特征，并使用LIFEx软件计算了其他特征。为了预测结果，我们实现了三个分类器-随机森林分类器，梯度增强决策树（GBDT）和决策树分类器。我们使用93个数据点训练了每个模型，并使用39个数据点测试了模型的性能。最佳模型-GBDT是根据性能指标选择的。

结果

观察到通常具有以下三个特征：MTV（新陈代谢肿瘤体积），原发肿瘤部位和GLCM_correlation对预测生存结果具有重要意义。对于测试队列，GBDT达到了88％的平衡准确度，而传统的统计模型报告了81.5％的平衡准确度。

结论

提出的分类器比现有技术具有更高的准确性。使用该分类器，我们可以估计HNSCC患者的结局，并可以根据结局选择治疗策略。