Predictive Value of Multiparametric MRI for Response to Single-Cycle Induction Chemo-Immunotherapy in Locally Advanced Head and Neck Squamous Cell Carcinoma.

OBJECTIVES

To assess the predictive value of multiparametric MRI for treatment response evaluation of induction chemo-immunotherapy in locally advanced head and neck squamous cell carcinoma.

METHODS

Twenty-two patients with locally advanced, histologically confirmed head and neck squamous cell carcinoma who were enrolled in the prospective multicenter phase II CheckRad-CD8 trial were included in the current analysis. In this unplanned secondary single-center analysis, all patients who received contrast-enhanced MRI at baseline and in week 4 after single-cycle induction therapy with cisplatin/docetaxel combined with the immune checkpoint inhibitors tremelimumab and durvalumab were included. In week 4, endoscopy with representative re-biopsy was performed to assess tumor response. All lesions were segmented in the baseline and restaging multiparametric MRI, including the primary tumor and lymph node metastases. The volume of interest of the respective lesions was volumetrically measured, and time-resolved mean intensities of the golden-angle radial sparse parallel-volume-interpolated gradient-echo perfusion (GRASP-VIBE) sequence were extracted. Additional quantitative parameters including the T1 ratio, short-TI inversion recovery ratio, apparent diffusion coefficient, and dynamic contrast-enhanced (DCE) values were measured. A model based on parallel random forests incorporating the MRI parameters from the baseline MRI was used to predict tumor response to therapy. Receiver operating characteristic (ROC) curves were used to evaluate the prognostic performance.

RESULTS

Fifteen patients (68.2%) showed pathologic complete response in the re-biopsy, while seven patients had a residual tumor (31.8%). In all patients, the MRI-based primary tumor volume was significantly lower after treatment. The baseline DCE parameters of time to peak and wash-out were significantly different between the pathologic complete response group and the residual tumor group (p < 0.05). The developed model, based on parallel random forests and DCE parameters, was able to predict therapy response with a sensitivity of 78.7% (95% CI 71.24-84.93) and a specificity of 78.6% (95% CI 67.13-87.48). The model had an area under the ROC curve of 0.866 (95% CI 0.819-0.914).

CONCLUSIONS

DCE parameters indicated treatment response at follow-up, and a random forest machine learning algorithm based on DCE parameters was able to predict treatment response to induction chemo-immunotherapy.