Click here to

Session: Emerging and Pre-clinical Imaging Technologies [Return to Session]

Boundary-Constrained Neural Network for Mouse Organ Segmentation

L Jiang*, Q Xu, A Chatziioannou, K Sheng, University of California Los Angeles, Los Angeles, CA


SU-J-207-3 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]

Room 207

Purpose: Intensity-modulated radiotherapy (IMRT) for mouse experiments is essential to mimic human treatments for clinical translation. A prerequisite for IMRT is a detailed delineation of involved organs. The lack of expertise to support mouse organ segmentation has hampered the utilization of IMRT. Automated mouse organ segmentation is difficult due to high noise and low organ contrast in micro-CT (µCT). We propose a boundary-constrained U-net (BCUnet) to improve the automated segmentation of major mouse organs.

Methods: The study is based on two datasets. The first dataset, including µCT mouse images with contrast (n=80) and without contrast (n=140) were provided by Rosenhain et al. The image resolution was 0.14mm×0.14mm×0.14mm. The second CT dataset including five mice was provided from an independent source. The image resolution is 0.2mm×0.2mm×0.2mm. 2D coronal CT slices with manual organ labels were input into a U-net. Two Sobel edge filters were used to extract ground truth and prediction boundary information before merging into an edge map, which was then fused as a binary probability map. The difference between ground truth and prediction is added to the loss function. The images were divided into training (n=160), validation (n=40), and testing (n=20). Additional testing was performed on the independent dataset. The performance of BCUnet was evaluated using Dice’s index (DSC) and average surface distance (ASD) and compared with the vanilla U-net.

Results: BCUnet achieved average DSC of 91.35%, 91.78%, 93.41%, 92.08%,91.55%, 91.48% and 86.32% for the bladder, lungs, heart, liver, intestine, kidneys, and spleen, respectively. BCUnet consistently outperformed the vanilla 2D U-net for all organs (P=0.0067). Except for the intestine, BCUnet improved the ASD of all organs. The robustness of BCUnet was also confirmed on the independent dataset.

Conclusion: By incorporating the boundary information in training, BCUnet improves the robustness and accuracy of automated organ segmentation for mouse µCT.


Segmentation, Image Processing, CT


IM/TH- image Segmentation: General (Most aspects)

Contact Email