Ensemble methods and partially-supervised learning for accurate and robust automatic murine organ segmentation

Abstract

Delineation of multiple organs in murine µCT images is crucial for preclinical studies but requires manual volumetric segmentation, a tedious and time-consuming process prone to inter-observer variability. Automatic deep learning-based segmentation can improve speed and reproducibility. While 2D and 3D deep learning models have been developed for anatomical segmentation, their generalization to external datasets has not been extensively investigated. Furthermore, ensemble learning, combining predictions of multiple 2D models, and partially-supervised learning (PSL), enabling training on partially-labeled datasets, have not been explored for preclinical purposes. This study demonstrates the first use of PSL frameworks and the superiority of 3D models in accuracy and generalizability to external datasets. Ensemble methods performed on par or better than the best individual 2D network, but only 3D models consistently generalized to external datasets (Dice Similarity Coefficient (DSC) > 0.8). PSL frameworks showed promising results across various datasets and organs, but its generalization to external data can be improved for some organs. This work highlights the superiority of 3D models over 2D and ensemble counterparts in accuracy and generalizability for murine µCT image segmentation. Additionally, a promising PSL framework is presented for leveraging multiple datasets without complete annotations. Our model can increase time-efficiency and improve reproducibility in preclinical radiotherapy workflows by circumventing manual contouring bottlenecks. Moreover, high segmentation accuracy of 3D models allows monitoring multiple organs over time using repeated µCT imaging, potentially reducing the number of mice sacrificed in studies, adhering to the 3R principle, specifically Reduction and Refinement.

Fig. 1

Image value distribution. The image value distribution shows better overlap after rescaling from (pseudo) Hounsfield units (HUs) to rescaled (pseudo) mass densities.

Fig. 2

Performance of the 2D (coronal, axial, sagittal), 2.5D (majority, soft, average voting) and 3D nnU-Net models. The performance on validation Datasets 4 & 5 and the Test Dataset was determined in terms of the DSC. The dotted gray line at DSC = 0.8 indicates the threshold for acceptable segmentations. The grey region indicates the 2.5D ensemble voting methods.

Fig. 3

Ground truth and 3D nnU-Net prediction for outlier case on Dataset 1. The ground truth labels for the lungs (blue and purple) showed a patchy appearance with parts of the lung missing.

Fig. 4

Influence of the number of CT projections on model performance. The 3D nnU-Net model shows increasing performance on Dataset 3, in terms of the Dice Similarity Coefficient (DSC), as the number of CT projections increases. For this dataset the cumulative exposure for different numbers of projections was 48, 96 and 192 mAs respectively.

Fig. 5

Partially-supervised learning cross-validation performance. Both partially-supervised learning frameworks, cnnU-Net and DoDNet, in terms of the DSC. For the lungs, heart, brain and bones, the different datasets are color-coded. For the spleen, the native and contrast-enhanced cases are color-coded.

Fig. 6

Partially supervised-learning test performance. Performance on the external Test Dataset of the partially-supervised learning models (cnnU-Net and DoDNet) compared to the dedicated single-organ networks, in terms of the DSC.

Fig. 7

Two representative segmentations by DoDNet of two different mice from the independent Test Dataset: (a) anterior (A) and posterior (P) view of a visually correct segmentation of both lungs, heart liver, kidneys, bladder, bones and brain, (b) anterior view of a visually incorrect bladder (dark green) and protruding liver (pink) segmentation.

Fig. 8

Spinal cord segmentation. The dedicated single-organ 3D nnU-Net model shows accurate spinal cord segmentation (in magenta).