Assessment of clinical radiosensitivity in patients with head-neck squamous cell carcinoma from pre-treatment quantitative ultrasound radiomics

Abstract

To investigate the role of quantitative ultrasound (QUS) radiomics to predict treatment response in patients with head and neck squamous cell carcinoma (HNSCC) treated with radical radiotherapy (RT). Five spectral parameters, 20 texture, and 80 texture-derivative features were extracted from the index lymph node before treatment. Response was assessed initially at 3 months with complete responders labelled as early responders (ER). Patients with residual disease were followed to classify them as either late responders (LR) or patients with persistent/progressive disease (PD). Machine learning classifiers with leave-one-out cross-validation was used for the development of a binary response-prediction radiomics model. A total of 59 patients were included in the study (22 ER, 29 LR, and 8 PD). A support vector machine (SVM) classifier led to the best performance with accuracy and area under curve (AUC) of 92% and 0.91, responsively to define the response at 3 months (ER vs. LR/PD). The 2-year recurrence-free survival for predicted-ER, LR, PD using an SVM-model was 91%, 78%, and 27%, respectively (p < 0.01). Pretreatment QUS-radiomics using texture derivatives in HNSCC can predict the response to RT with an accuracy of more than 90% with a strong influence on the survival.Clinical trial registration: clinicaltrials.gov.in identifier NCT03908684.

Figure 1

Representative ultrasound B-mode (uppermost row), QUS spectral parametric maps of AAC and MBF, and the corresponding texture images from one patient in each of the three response groups-early responder (a), late responder (b), and persistent disease (c). QUS parametric images include the largest involved lymph node (central region bounded by closed dotted white curve). The colour bar ranges are 0 to 150 dB/cm³ for AAC, − 20 to 20 dB for MBF and arbitrary unit for the texture features. The scale bar represents 1 cm.

Figure 2

ROC plots of predictive models for different endpoints developed using spectral and texture features alone (upper row) and those developed using all the spectral, texture, and texture-derivate features (lower row). The endpoints considered are 3-month complete responder versus partial responder/non-responder (a,c), late responder versus persistent/progressive disease (b,d). Three standard classification algorithms that include FLD, KNN, and SVM-RBF were evaluated as indicated in the inset legend. The classification models that include texture-derivate features (lower row) achieved higher AUC values in general compared to those developed without texture-derivate features.

Figure 3

Hyperplane plot with the decision boundary based on support vector machine classifier using three features to differentiate the complete responders from partial/non-responders at three months following radiotherapy completion.

Figure 4

Kaplan Meier survival plots showing the recurrence-free survival for the three predicted groups using a support vector machine radiomics model.

Figure 5

Flowchart showing the study methodology.