Strategies for tackling the class imbalance problem of oropharyngeal primary tumor segmentation on magnetic resonance imaging. (July 2022)

• Record Type: Journal Article
• Authors: Rodríguez Outeiral, Roque; Bos, Paula; van der Hulst, Hedda J.; Al-Mamgani, Abrahim; Jasperse, Bas; Simões, Rita; van der Heide, Uulke A.
• Is Part Of: Physics and imaging in radiation oncology. Volume 23(2022)
• Page Start: 144
• Page End: 149
• Publication Date: 2022-07
• Subjects: Oropharyngeal cancer -- Convolutional neural network -- Segmentation -- Class imbalance, MRI -- Two-stage approach; Radiotherapy -- Periodicals; Radiation dosimetry -- Periodicals; Cancer -- Imaging -- Periodicals; Oncology -- Periodicals; 615.842
• DOI: 10.1016/j.phro.2022.08.005
• Languages: English
• ISSNs: 2405-6316

Abstract

Highlights: Loss functions that account for class imbalance may not improve the segmentations. Our two-stage segmentation approach can outperform the 3D U-Net. A fully-automatic two-stage approach can perform comparably to a semi-automatic approach. Abstract: 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.