Oropharyngeal primary tumor segmentation for radiotherapy planning on magnetic resonance imaging using deep learning

Abstract

Background and purpose: Segmentation of oropharyngeal squamous cell carcinoma (OPSCC) is needed for radiotherapy planning. We aimed to segment the primary tumor for OPSCC on MRI using convolutional neural networks (CNNs). We investigated the effect of multiple MRI sequences as input and we proposed a semiautomatic approach for tumor segmentation that is expected to save time in the clinic. Materials and methods: We included 171 OPSCC patients retrospectively from 2010 until 2015. For all patients the following MRI sequences were available: T1-weighted, T2-weighted and 3D T1-weighted after gadolinium injection. We trained a 3D UNet using the entire images and images with reduced context, considering only information within clipboxes around the tumor. We compared the performance using different combinations of MRI sequences as input. Finally, a semi-automatic approach by two human observers defining clipboxes around the tumor was tested. Segmentation performance was measured with Sørensen-Dice coefficient (Dice), 95th Hausdorff distance (HD) and Mean Surface Distance (MSD). Results: The 3D UNet trained with full context and all sequences as input yielded a median Dice of 0.55, HD of 8.7 mm and MSD of 2.7 mm. Combining all MRI sequences was better than using single sequences. The semiautomatic approach with all sequences as input yielded significantly better performance (p < 0.001): a median Dice of 0.74, HD of 4.6 mm and MSD of 1.2 mm. Conclusion:Reducing the amount of context around the tumor and combining multiple MRI sequences improved the segmentation performance. A semi-automatic approach was accurate and clinically feasible.

Evaluation of deep learning-based multiparametric MRI oropharyngeal primary tumor auto-segmentation and investigation of input channel effects: Results from a prospective imaging registry

“…DL = deep learning, N = number of images sets used in study, GTVp = primary gross tumor volume, DSC = Dice similarity coefficient, OPC = oropharyngeal cancer, NPC = nasopharyngeal cancer, HNSCC = head and neck squamous cell carcinoma, T2 = T2-weighted MRI, T1 = T1-weighted MRI, DCE = dynamic contrast enhanced MRI, DWI = diffusion weighted imaging MRI, LOOCV = leave-one-out cross-validation, CV = cross-validation, CNN = convolutional neural network. Author, Year Site Modality DL Architecture N (Train, Test) GTVp DSC (average) This study, 2021 OPC MRI (T1, T2, DCE, DWI) 3D Residual Unet 30 (LOOCV) 0.73 (best model, T2 + T1) Outeiral et al, 2021 [ 27 ] OPC MRI (T1, T2) 3D Unet 171 (151, 20) 0.55 Andrearczyk et al, 2020 [ 32 ] OPC CT, PET 2D Unet 202 (LOOCV) 0.48 (CT), 0.58 (PET), 0.6 (CT + PET) Moe et al, 2019 [ 33 ] OPC CT, PET 2D Unet 197 (157, 40) 0.65 (CT), 0.71 (PET), 0.75 (CT + PET) Naser et al, 2020 [ 35 ] OPC CT, PET 3D Unet 201 (5-fold CV) 0.69 Iantsen et al, 2020 [ 36 ] OPC CT, PET 3D Unet 201 (4-fold CV) 0.75 Ma et al, 2018 [ 19 ] NPC MRI (T1) 3D CNN + graph-cut 30 (LOOCV) 0.85 Ye et al, 2020 [ 18 ] NPC MRI (T1, T2) 3D Unet 44 (10-fold CV) 0.62 (T1), 0.64 (T2), 0.72 (T1 + T2) Chen et al, 2020 [ 21 ] NPC MRI (T1, T2) 3D Encoder-decoder network 149 (5-fold CV) 0.72 Huang et al, 2019 [ 26 ] …”

“…Notably, compared to previous fully-automated primary tumor segmentation studies of HNSCC patients, we achieved promising average DSC performance ( Table 2 ). While it is difficult to directly compare DSCs between studies due to different datasets and model training, our models seemingly improve upon the only other fully-automated OPC tumor segmentation study to our knowledge, which exclusively investigated anatomical MRI [ 27 ].…”

“…Importantly, we recently demonstrated the utility of DL for HNSCC organ-at-risk auto-segmentation using MRI, with improvements in performance, execution time, and dosimetric differences compared to other auto-segmentation methods [ 16 ]. While several DL tumor auto-segmentation studies for nasopharyngeal cancer using MRI have been published [ [17] , [18] , [19] , [20] , [21] , [22] , [23] , [24] , [25] , [26] ], to our knowledge, only one study has been published for OPC [ 27 ]. Since HNSCC tumors at different anatomical sites have distinct anatomic boundaries and characteristics [ [28] , [29] ], it is crucial that tumor segmentation models are developed for each site accordingly.…”

Challenges and chances for deep-learning based target and organ at risk segmentation in radiotherapy of head and neck cancer

“…The first was to implement a fully automated 2-stage approach, where first a UNet was trained to localize a bounding box around the GTV, followed by training a UNet to segment the GTV using cropped data. The authors earlier published on the use of observer defined bounding boxes to improve segmentation results in a subset of this dataset [14] . The second approach was to compare four different loss functions, Dice-loss, Generalized-Dice-loss, Tversky-loss, and Unified-Focal-loss.…”

Intensity standardization methods in magnetic resonance imaging of head and neck cancer

“…Notably, the increasing use of MRI-guided technology for adaptive HNC radiotherapy will likely increase the clinical integration of MRI quantitative analysis [12] . While several recent HNC studies have implemented cohort-level quantitative analysis of conventional weighted MRI [13] , [14] , [15] , [16] , [17] , [18] , [19] , [20] , [21] , [22] , relatively few have investigated incorporating intensity standardization into processing pipelines [16] , [17] , [18] , [19] , [20] , [22] , and even fewer have tested multiple standardization methods [20] . Furthermore, while rigorous studies have tested MRI intensity standardization methods for various anatomical regions, chiefly the brain [5] , [23] , such methods for the head and neck region have yet to be systematically investigated.…”