A network score-based metric to optimize the quality assurance of automatic radiotherapy target segmentations.

Background and purpose: Existing methods for quality assurance of the radiotherapy auto-segmentations focus on the correlation between the average model entropy and the Dice Similarity Coefficient (DSC) only. We identified a metric directly derived from the output of the network and correlated it with clinically relevant metrics for contour accuracy. Materials and Methods: Magnetic Resonance Imaging auto-segmentations were available for the gross tumor volume for cervical cancer brachytherapy (106 segmentations) and for the clinical target volume for rectal cancer external-beam radiotherapy (77 segmentations). The nnU-Net's output before binarization was taken as a score map. We defined a metric as the mean of the voxels in the score map above a threshold (λ). Comparisons were made with the mean and standard deviation over the score map and with the mean over the entropy map. The DSC, the 95th Hausdorff distance, the mean surface distance (MSD) and the surface DSC were computed for segmentation quality. Correlations between the studied metrics and model quality were assessed with the Pearson correlation coefficient (r). The area under the curve (AUC) was determined for detecting segmentations that require reviewing. Results: For both tasks, our metric (λ = 0.30) correlated more strongly with the segmentation quality than the mean over the entropy map (for surface DSC, r > 0.65 vs. r < 0.60). The AUC was above 0.84 for detecting MSD values above 2 mm. Conclusions: Our metric correlated strongly with clinically relevant segmentation metrics and detected segmentations that required reviewing, indicating its potential for automatic quality assurance of radiotherapy target auto-segmentations.

Deep learning for segmentation of the cervical cancer gross tumor volume on magnetic resonance imaging for brachytherapy.

BACKGROUND: Segmentation of the Gross Tumor Volume (GTV) is a crucial step in the brachytherapy (BT) treatment planning workflow. Currently, radiation oncologists segment the GTV manually, which is time-consuming. The time pressure is particularly critical for BT because during the segmentation process the patient waits immobilized in bed with the applicator in place. Automatic segmentation algorithms can potentially reduce both the clinical workload and the patient burden. Although deep learning based automatic segmentation algorithms have been extensively developed for organs at risk, automatic segmentation of the targets is less common. The aim of this study was to automatically segment the cervical cancer GTV on BT MRI images using a state-of-the-art automatic segmentation framework and assess its performance. METHODS: A cohort of 195 cervical cancer patients treated between August 2012 and December 2021 was retrospectively collected. A total of 524 separate BT fractions were included and the axial T2-weighted (T2w) MRI sequence was used for this project. The 3D nnU-Net was used as the automatic segmentation framework. The automatic segmentations were compared with the manual segmentations used for clinical practice with Sørensen-Dice coefficient (Dice), 95th Hausdorff distance (95th HD) and mean surface distance (MSD). The dosimetric impact was defined as the difference in D98 (ΔD98) and D90 (ΔD90) between the manual segmentations and the automatic segmentations, evaluated using the clinical dose distribution. The performance of the network was also compared separately depending on FIGO stage and on GTV volume. RESULTS: The network achieved a median Dice of 0.73 (interquartile range (IQR) = 0.50-0.80), median 95th HD of 6.8 mm (IQR = 4.2-12.5 mm) and median MSD of 1.4 mm (IQR = 0.90-2.8 mm). The median ΔD90 and ΔD98 were 0.18 Gy (IQR = -1.38-1.19 Gy) and 0.20 Gy (IQR =-1.10-0.95 Gy) respectively. No significant differences in geometric or dosimetric performance were observed between tumors with different FIGO stages, however significantly improved Dice and dosimetric performance was found for larger tumors. CONCLUSIONS: The nnU-Net framework achieved state-of-the-art performance in the segmentation of the cervical cancer GTV on BT MRI images. Reasonable median performance was achieved geometrically and dosimetrically but with high variability among patients.

Strategies for tackling the class imbalance problem of oropharyngeal primary tumor segmentation on magnetic resonance imaging.

Background and purpose: Contouring oropharyngeal primary tumors in radiotherapy is currently done manually which is time-consuming. Autocontouring techniques based on deep learning methods are a desirable alternative, but these methods can render suboptimal results when the structure to segment is considerably smaller than the rest of the image. The purpose of this work was to investigate different strategies to tackle the class imbalance problem in this tumor site. Materials and methods: A cohort of 230 oropharyngeal cancer patients treated between 2010 and 2018 was retrospectively collected. The following magnetic resonance imaging (MRI) sequences were available: T1-weighted, T2-weighted, 3D T1-weighted after gadolinium injection. Two strategies to tackle the class imbalance problem were studied: training with different loss functions (namely: Dice loss, Generalized Dice loss, Focal Tversky loss and Unified Focal loss) and implementing a two-stage approach (i.e. splitting the task in detection and segmentation). Segmentation performance was measured with Sørensen-Dice coefficient (Dice), 95th Hausdorff distance (HD) and Mean Surface Distance (MSD). Results: The network trained with the Generalized Dice Loss yielded a median Dice of 0.54, median 95th HD of 10.6 mm and median MSD of 2.4 mm but no significant differences were observed among the different loss functions (p-value > 0.7). The two-stage approach resulted in a median Dice of 0.64, median HD of 8.7 mm and median MSD of 2.1 mm, significantly outperforming the end-to-end 3D U-Net (p-value < 0.05). Conclusion: No significant differences were observed when training with different loss functions. The two-stage approach outperformed the end-to-end 3D U-Net.

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

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 semi-automatic 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 semi-automatic 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.