Segmentation of regional brain volumes from CT images with 3D deep learning models
Poster No:
1899
Submission Type:
Abstract Submission
Authors:
Aram Salehi1, Natalia Vilor-Tejedor2, Juan Domingo Gispert2, Hieab Adams1, Tavia Evans3
Institutions:
1Department of Human Genetics, Radboud UMC, Nijmegen, Netherlands, 2BarcelonaBeta Brain Research Center, Barcelona, Catalonia, 3Erasmus University Medical Center, Rotterdam, Zuid Holland
First Author:
Co-Author(s):
Introduction:
Brain imaging is vital for studying the human brain and unraveling key insights into the onset, progression, and prognosis of brain diseases. Magnetic resonance imaging (MRI) provides high detail and soft tissue contrast, however is unfeasible in certain patient populations and many regions lack proper access. Computerized tomography (CT) offers cheaper, faster, and more accessible alternative, but its lower soft tissue contrast makes effective segmentation challenging. In this study, we address this challenge by evaluating deep learning-based models to enhance the accuracy of brain segmentation in CT scans by detecting intricate patterns that may be overlooked by traditional segmentation methods.
Methods:
We used 356 paired CT/MRI brain images from the ALFA (ALzheimer and FAmilies) project[4]. Segmentations of regional volumes from MRI images were extracted using FreeSurfer and used as the 'standard of truth' for model training[2]. CT images were aligned to the MRI images and the segmentation labels (FSL, FLIRT)[5]. The CT image voxel intensities were restricted to reflect the brain window of Hounsfield units (0-80). The dataset was split into 70%, 15%, and 15% for training, validation, and testing respectively. We investigated three base 3D deep learning models known to perform well within image segmentation; SwinUNETER, HighResNet, and HRNet [1,3,6]. These networks were trained using CT brain images and paired MRI-derived labels to extract 29 distinct regions. The training of models involved utilizing a combination of the dice score and categorical cross-entropy function as the loss function. The evaluation of segmentation performance was conducted using dice similarity metrics.
Results:
SwinUNETR outperformed other models, achieving a mean dice similarity score of 82.71±5.2 ( followed by HRNet: 81.08±5.09 and HighResNet: 79.99±5.78) across the 29 regions. The lowest mean dice similarity, 71.49, was observed for the right cerebral cortex (HRNet: 72.85, HighResNet: 68.53), while the mean highest score of 91.58 was achieved for the brain stem (HRNet: right thalamus 89.1, HighResNet: brain-stem 91.72) across the 54 test images (Figures 1 & 2). Our results indicate that SwinUNETR tends to overestimate regional volume segmentations, with a mean deviation of 1.04 and a standard deviation of 0.06 when considering normalized percentage data of the regional volumes. The results demonstrated that for all models larger regions, such as the cerebral cortex (gray matter) and cerebral white matter, displayed a lower mean dice similarity metric, while more internal regions like the thalamus and lateral ventricle exhibited superior performance (Figure 1).
·Figure 1. Left; Dice similarity index for HighReNet, SwinUNETR and HRNet, Right; Percentage of the total segmentation for each region.
·Figure 2. Best performing test case with regards to application of the model. Left to right; Original CT image, predicted segmentation using SwinUNETR, and ground truth from FreeSurfer MRI processing.
Conclusions:
Here we show that regional brain volumes can be reliably segmented from CT images using deep learning models. The results showed that SwinUNETER performed the best overall for segmentation of brain CT scans particularly for various subcortical structures. The lowest concordance with MRI segmentations was seen for the cerebral grey and white matter, pointing to the potential inherent difficulty of differentiating between the two because of the low tissue contrast from CT imaging. Future studies could assess whether other models can better distinguish between grey and white matter segmentation, and whether CT images could be used to generate other imaging markers beyond regional volumetry. In conclusion, our approach shows promise for extracting clinically relevant information from CT scans and could be crucial for unlocking the value of CT images in resource-constrained settings, where MRI is not readily available.
Modeling and Analysis Methods:
Methods Development 1
Other Methods
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Anatomy and Brain Mapping 2
Novel Imaging Acquisition Methods:
Anatomical MRI
Imaging Methods Other
Keywords:
Computed Tomography (CT)
Data analysis
Machine Learning
MRI
Other - Brain Segmentation
1|2Indicates the priority used for review
Provide references using author date format
2-Iglesias, J. E.(2015). A computational atlas of the hippocampal formation using ex vivo, ultra-high resolution MRI: application to adaptive segmentation of in vivo MRI. Neuroimage 115, 117-137.
3-Li W.(2017). On the compactness, efficiency, and representation of 3D convolutional networks: brain parcellation as a pretext task. InInformation Processing in Medical Imaging: 25th International Conference, IPMI 2017, Boone, NC, USA, Proceedings (pp. 348-360). Springer International Publishing.
4-Molinuevo JL.(2016) The ALFA project: a research platform to identify early pathophysiological features of Alzheimer's disease. Alzheimer's & Dementia: Translational Research & Clinical Interventions ;2(2):82-92.
5-Patenaude, B.(2011) A Bayesian model of shape and appearance for subcortical brain segmentation. Neuroimage 56, 907-922.
6-Wang J. (2020).Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence. ;43(10):3349-64.