Robustness of brain tumor segmentation

Deep learning for automatic head and neck lymph node level delineation provides expert-level accuracy

Background

Deep learning-based head and neck lymph node level (HN_LNL) autodelineation is of high relevance to radiotherapy research and clinical treatment planning but still underinvestigated in academic literature. In particular, there is no publicly available open-source solution for large-scale autosegmentation of HN_LNL in the research setting.

Methods

An expert-delineated cohort of 35 planning CTs was used for training of an nnU-net 3D-fullres/2D-ensemble model for autosegmentation of 20 different HN_LNL. A second cohort acquired at the same institution later in time served as the test set (n = 20). In a completely blinded evaluation, 3 clinical experts rated the quality of deep learning autosegmentations in a head-to-head comparison with expert-created contours. For a subgroup of 10 cases, intraobserver variability was compared to the average deep learning autosegmentation accuracy on the original and recontoured set of expert segmentations. A postprocessing step to adjust craniocaudal boundaries of level autosegmentations to the CT slice plane was introduced and the effect of autocontour consistency with CT slice plane orientation on geometric accuracy and expert rating was investigated.

Results

Blinded expert ratings for deep learning segmentations and expert-created contours were not significantly different. Deep learning segmentations with slice plane adjustment were rated numerically higher (mean, 81.0 vs. 79.6, p = 0.185) and deep learning segmentations without slice plane adjustment were rated numerically lower (77.2 vs. 79.6, p = 0.167) than manually drawn contours. In a head-to-head comparison, deep learning segmentations with CT slice plane adjustment were rated significantly better than deep learning contours without slice plane adjustment (81.0 vs. 77.2, p = 0.004). Geometric accuracy of deep learning segmentations was not different from intraobserver variability (mean Dice per level, 0.76 vs. 0.77, p = 0.307). Clinical significance of contour consistency with CT slice plane orientation was not represented by geometric accuracy metrics (volumetric Dice, 0.78 vs. 0.78, p = 0.703).

Conclusions

We show that a nnU-net 3D-fullres/2D-ensemble model can be used for highly accurate autodelineation of HN_LNL using only a limited training dataset that is ideally suited for large-scale standardized autodelineation of HN_LNL in the research setting. Geometric accuracy metrics are only an imperfect surrogate for blinded expert rating.

Region-based evidential deep learning to quantify uncertainty and improve robustness of brain tumor segmentation

Despite recent advances in the accuracy of brain tumor segmentation, the results still suffer from low reliability and robustness. Uncertainty estimation is an efficient solution to this problem, as it provides a measure of confidence in the segmentation results. The current uncertainty estimation methods based on quantile regression, Bayesian neural network, ensemble, and Monte Carlo dropout are limited by their high computational cost and inconsistency. In order to overcome these challenges, Evidential Deep Learning (EDL) was developed in recent work but primarily for natural image classification and showed inferior segmentation results. In this paper, we proposed a region-based EDL segmentation framework that can generate reliable uncertainty maps and accurate segmentation results, which is robust to noise and image corruption. We used the Theory of Evidence to interpret the output of a neural network as evidence values gathered from input features. Following Subjective Logic, evidence was parameterized as a Dirichlet distribution, and predicted probabilities were treated as subjective opinions. To evaluate the performance of our model on segmentation and uncertainty estimation, we conducted quantitative and qualitative experiments on the BraTS 2020 dataset. The results demonstrated the top performance of the proposed method in quantifying segmentation uncertainty and robustly segmenting tumors. Furthermore, our proposed new framework maintained the advantages of low computational cost and easy implementation and showed the potential for clinical application.

Asymmetric Ensemble of Asymmetric U-Net Models for Brain Tumor Segmentation With Uncertainty Estimation

Accurate brain tumor segmentation is crucial for clinical assessment, follow-up, and subsequent treatment of gliomas. While convolutional neural networks (CNN) have become state of the art in this task, most proposed models either use 2D architectures ignoring 3D contextual information or 3D models requiring large memory capacity and extensive learning databases. In this study, an ensemble of two kinds of U-Net-like models based on both 3D and 2.5D convolutions is proposed to segment multimodal magnetic resonance images (MRI). The 3D model uses concatenated data in a modified U-Net architecture. In contrast, the 2.5D model is based on a multi-input strategy to extract low-level features from each modality independently and on a new 2.5D Multi-View Inception block that aims to merge features from different views of a 3D image aggregating multi-scale features. The Asymmetric Ensemble of Asymmetric U-Net (AE AU-Net) based on both is designed to find a balance between increasing multi-scale and 3D contextual information extraction and keeping memory consumption low. Experiments on 2019 dataset show that our model improves enhancing tumor sub-region segmentation. Overall, performance is comparable with state-of-the-art results, although with less learning data or memory requirements. In addition, we provide voxel-wise and structure-wise uncertainties of the segmentation results, and we have established qualitative and quantitative relationships between uncertainty and prediction errors. Dice similarity coefficient for the whole tumor, tumor core, and tumor enhancing regions on BraTS 2019 validation dataset were 0.902, 0.815, and 0.773. We also applied our method in BraTS 2018 with corresponding Dice score values of 0.908, 0.838, and 0.800.