Auto-Segmentation of Oropharyngeal Cancer Primary Tumors Using Multiparametric MRI-Based Deep Learning

Abstract

<h3>Purpose/Objective(s)</h3> Segmentation of oropharyngeal cancer (OPC) primary gross tumor volumes (GTVp) is crucial for radiotherapy planning. Multiparametric MRI (mpMRI) acquisitions that incorporate anatomical and functional sequences are being increasingly used for OPC-adaptive radiotherapy workflows and will play a prominent role in GTVp segmentation. However, human-generated segmentations are error-prone and suffer from high inter-observer variability. Therefore, we constructed mpMRI deep learning OPC GTVp auto-segmentation models and determined the impact of variable mpMRI channels on segmentation performance. <h3>Materials/Methods</h3> GTVp ground truth segmentations were manually generated for 30 OPC patients on T2-weighted (T2) MRI from a prospective clinical trial. We evaluated 5 mpMRI input channels (T2, T1-weighted [T1], apparent diffusion coefficient [ADC], volume transfer constant [Ktrans], and fractional volume [Ve]). 3D Residual U-net models for different channel combinations were developed and assessed using a leave-one-out cross-validation approach. A baseline T2 model was compared to mpMRI models (T2+T1, T2+ADC, T2+Ktrans, T2+Ve, all five channels [ALL]) primarily using the Dice similarity coefficient (DSC). Sensitivity, positive predictive value, Hausdorff distance (HD), false-negative DSC (FND), false-positive DSC, surface DSC, 95% HD, and mean surface distance were also assessed. For the best model, ground truth and model-generated segmentations were compared through a Turing test using three physician observers. <h3>Results</h3> Our models demonstrated high performance across all evaluation metrics, with mean DSCs of 0.71 (ALL) to 0.73 (T2+T1). Compared to the baseline T2 model, performance was significantly improved for HD, FND, sensitivity, Surface DSC, and 95% HD for the T2+T1 model (p<0.05) and for FND for the T2+Ve and ALL models (p<0.05). No model demonstrated significant correlations between tumor size and DSC (p>0.05). Most models demonstrated significant correlations between tumor size and HD or Surface DSC (p<0.05), with the exception of those that included ADC or Ve as input channels (p>0.05). The Turing test revealed no significant difference between the best performing model (T2+T1) and the ground truth segmentations for all expert observers (p>0.05). <h3>Conclusion</h3> Deep learning using mpMRI data provides high-quality segmentations of OPC GTVp, where clinical experts are unable to determine any meaningful differences between ground truth and model segmentations. The incorporation of additional mpMRI channels may increase the performance of certain evaluation metrics and improve model robustness and consistency. This pilot study is a promising step towards future large-scale studies for fully automated MR-guided OPC radiotherapy.