[go: up one dir, main page]
More Web Proxy on the site http://driver.im/ Skip to main content
PMC home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2023 Nov 27.
Published in final edited form as: Proc SPIE Int Soc Opt Eng. 2023 Apr 3;12464:124642G. doi: 10.1117/12.2653999

Investigation of probability maps in deep-learning-based brain ventricle parcellation

Yuli Wang a, Anqi Feng a, Yuan Xue b, Muhan Shao b, Ari M Blitz c, Mark G Luciano d, Aaron Carass b, Jerry L Prince b
PMCID: PMC10679955  NIHMSID: NIHMS1945636  PMID: 38013746

Abstract

Normal Pressure Hydrocephalus (NPH) is a brain disorder associated with ventriculomegaly. Accurate segmentation of the ventricle system into its sub-compartments from magnetic resonance images (MRIs) could help evaluate NPH patients for surgical intervention. In this paper, we modify a 3D U-net utilizing probability maps to perform accurate ventricle parcellation, even with grossly enlarged ventricles and post-surgery shunt artifacts, from MRIs. Our method achieves a mean dice similarity coefficient (DSC) on whole ventricles for healthy controls of 0.864 ± 0.047 and 0.961 ± 0.024 for NPH patients. Furthermore, with the benefit of probability maps, the proposed method provides superior performance on MRI with grossly enlarged ventricles (mean DSC value of 0.965 ± 0.027) or post-surgery shunt artifacts (mean DSC value of 0.964 ± 0.031). Results indicate that our method provides a high robust parcellation tool on the ventricular systems which is comparable to other state-of-the-art methods.

Keywords: normal pressure hydrocephalus, ventricle parcellation, MRI, probability map

1. INTRODUCTION

Normal pressure hydrocephalus (NPH) is a brain disorder that may cause incontinence, gait unsteadiness, and dementia in the elderly;1 it can be treated with shunt placement to drain excess cerebrospinal fluid (CSF).2 NPH is characterized by enlargement of the brain ventricles while maintaining normal CSF pressure levels. Its diagnosis is based on brain imaging and clinical symptoms.3 Since disproportionate ventricular enlargement could impact diagnosis and treatment,3 accurate segmentation into four ventricular cavities is clinically important.

Previous ventricle segmentation methods4–16 include FreeSurfer13 and multi-atlas methods,7 which can take hours to run and has decreased performance on enlarged ventricle subjects. RUDOLPH6 was specifically designed to segment subjects with enlarged ventricles but also has a multi-hour run time. Since 2012, deep convolutional networks have shown great advantages in image segmentation tasks17–21 and caught up with humans in classification performance. One recent algorithm, VParNet,9 utilizes a 3D CNN with a U-net architecture, achieved accurate results on moderate ventricular enlargement subjects and takes approximately 2 minutes of CPU time. However, due to its limitations, VParNet cannot handle severely enlarged ventricles or ventricles with post-shunt artifacts.

In this paper, we present a 3D CNN with the U-Net architecture and residual blocks for brain ventricle segmentation in magnetic resonance images (MRIs). We incorporate probability maps to improve network robustness. We conducted experiments over three cohorts: healthy controls, NPH patients, and NPH patients post surgical intervention with shunt placement. Our method produced results that are significantly better than the state-of-the-art alternatives and was successful with data that has shunt artifacts.

METHOD

Data sets and Pre-processing

T1-weighted (T1w) magnetically prepared rapidly acquired gradient echo (MPRAGE) images from three cohorts were used to train or evaluate our method. All three cohorts were manually delineated with four labels: left and right lateral ventricles, third ventricle, and fourth ventricle. The first cohort comprised 50 images from Neuromorphometrics Inc. (NMM). Each image was manually delineated with five labels: left and right lateral ventricles, third ventricle, fourth ventricle, and a catch-all background label. 15 images were used for training, 5 were used for validation, and the remaining 30 were used for testing. The second cohort was from an NPH database that includes 95 NPH patients with images from a 3T MR scanner (Siemens Corporation, Germany). 25 images were used for training, 5 for validation, and the remaining 65 were used for testing. The third cohort consists of six images from post-surgery shunt implant NPH patients (PS-NPH). These images have shunt artifacts, shown in Fig. 4(d). These six images were only used for network evaluation. All images were run though a pre-processing pipeline: N4 bias correction,22 rigid registration to an MNI-space, skull stripping,23 and intensity normalization.24

Figure 4.

Figure 4.

Open in a new tab

Segmentation results from our method (Ours) and manual delineations (Manual) on the (a) NMM and (b) NPH data sets. Visual comparison among FreeSurfer, VParNet, and our method (Ours) for (c) a severely enlarged ventricle subject and (d) a subject with a post-surgery shunt.

Algorithm overview and segmentation network

We trained our 3D U-net with the pre-processed data as well as ground truth delineations and label-wise probability maps created from the delineations (details below). Figure 1 illustrates the algorithm’s main components. The network includes a series of contracting encoder blocks that extract image features from local to global contexts and a series of decoder blocks with skip connections from the encoders and shortcuts within the decoders which act as “learnable” up-samplers.

Figure 1.

Figure 1.

Open in a new tab

Flowchart for MRI ventricle segmentation. Segmentation of the MRIs is achieved with a modified 3D U-Net. The input data to the 3D U-Net is pre-processed from the original data. The loss function used to train the network consists of the DSC loss of all labels between ground truth delineation data and network output data, as well as the binary cross-entropy loss between the probability map and network output data.

Implementation

The regional probability maps are generated by averaging all 25 ground truth annotations from the NPH cohort in MNI-space. The loss function used to train the network is given as

Loss=(1−λ)(1−1L∑l=1Lε+2∑iPilTilε+∑iPil+∑iTil)+λ(1L∑l=1L∑i(Pillog(Gil))+(1−Pil)log(1−Gil)), (1)

where l ∈ (1,…,L) is the label index, L is the total number of labels. λ is a hyperparameter (= 0.1), Pil is the probability that voxel i has label l generated by the network, Til is the binary value indicating if voxel i should be labeled l, and Gil is the value of the probability map at voxel i for label l. ε (= 10−3) is used to avoid a zero denominator during training. Data augmentation during training includes left-right flipping, elastic deformation, and rotation. Our network was trained for 150 epochs using the Adam optimizer (step size of α = 10−3).

3. EXPERIMENTS AND RESULTS

In the first experiment, we ran FreeSurfer, VParNet, and our method on the 30 testing subjects from the NMM cohort and 65 subjects from our NPH cohort. We computed the DSC for the two datasets and report the results in Fig. 3(a) and (b). The DSC value of our method is 0.864 ± 0.047 on NMM and 0.961 ± 0.024 on NPH, which achieve comparable performance with VParNet. We break down the 65 NPH subject into 55 enlarged ventricle subjects and ten severely enlarged ventricle subjects and run the aforementioned three segmentation methods on these ten subjects with severe ventricle enlargement. Our method achieves DSC of 0.965 ± 0.027 on NPH with severe enlargement as shown in Fig. 3(c). A visualization of the ventricle parcellation produced by our method is shown in Figures 4(a) and (b) for NMM and NPH data, respectively. Figure 4(c) shows failure cases for FreeSurfer and VParNet on an NPH subject with severely enlarged ventricles.

Figure 3.

Figure 3.

Open in a new tab

Box plots of the DSC: (a) 30 test subjects from NMM data set, (b) 65 test subjects NPH dataset, (c) an additional 10 test subjects from NPH dataset with severe ventricle enlargement, and (d) our six test subjects from the PS-NPH dataset. Key: RLV - Right lateral ventricle; LLV - left lateral ventricle; 3rd - third ventricle; 4th - fourth ventricle; and Whole - whole ventricular system.

In our second experiment, we ran FreeSurfer, VParNet, and our method on our six PS-NPH subjects. FreeSurfer and VParNet, again, had failure cases as illustrated in Fig. 4(d). The DSC value comparison for the PS-NPH cohort are summarized in Fig. 3(d), where our method achieves a DSC of 0.964 ± 0.031.

4. DISCUSSION AND CONCLUSION

We presented a 3D U-net network utilizing probability maps to segment the ventricle system in patients with normal, enlarged ventricles, severely enlarged ventricles, and with post-surgery shunt artifacts. The model was trained on two different datasets and was evaluated on 30 NMM subjects, 55 NPH subjects, 10 NPH subjects with severely enlarged ventricles, and 6 post-surgery shunt NPH patients. Overall, our method achieved comparable or better results than the state-of-the-art methods. Our proposed method was the only method that could handle severe ventricle enlargement and patients with post-surgery shunts; demonstrating the robustness introduced by the inclusion of our probability maps in our network.

Figure 2.

Figure 2.

Open in a new tab

(a) Architecture of proposed ventricles parcellation network. The shape of tensor is denoted at each resolution level. (b) Architecture of the encoder block. (c) Architecture of the decoder block.

Acknowledgments

This work was supported by the NIH / NINDS under grant U01-NS122764 (PI: M.G. Luciano).

REFERENCES

  • [1].Jaraj D, Rabiei K, Marlow T, Jensen C, Skoog I, and Wikkelsø C, “Prevalence of idiopathic normal-pressure hydrocephalus,” Neurology 82(16), 1449–1454 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [2].Shprecher D, Schwalb J, and Kurlan R, “Normal pressure hydrocephalus: Diagnosis and treatment,” Current Neurology and Neuroscience Reports 8(5), 371–376 (2008). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [3].Ishikawa M, Hashimoto M, Kuwana N, Mori E, Miyake H, Wachi A, Takeuchi T, Kazui H, and Koyama H, “Guidelines for Management of Idiopathic Normal Pressure Hydrocephalus Guidelines From the Guidelines Committee of Idiopathic Normal Pressure Hydrocephalus, the Japanese Society of Normal Pressure Hydrocephalus,” Neurologia Medico-chirurgica 48(Supplement), S1–S23 (2008). [DOI] [PubMed] [Google Scholar]
  • [4].Shiee N, Bazin P-L, Cuzzocreo JL, Blitz A, and Pham DL, “Segmentation of brain images using adaptive atlases with application to ventriculomegaly,” in [22nd Inf. Proc. in Med. Imaging (IPMI 2011)], Lecture Notes in Computer Science 6801, 1–12, Springer; Berlin Heidelberg: (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [5].Ellingsen LM, Roy S, Carass A, Blitz AM, Pham DL, and Prince JL, “Segmentation and labeling of the ventricular system in normal pressure hydrocephalus using patch-based tissue classification and multi-atlas labeling,” in [Proceedings of SPIE Medical Imaging (SPIE-MI 2016), San Diego, CA, February 27 – March 3, 2016], 9784, 97840G–97840G–7 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [6].Carass A, Shao M, Li X, Dewey BE, Blitz AM, Roy S, Pham DL, Prince JL, and Ellingsen LM, “Whole brain parcellation with pathology: Validation on ventriculomegaly patients,” in [International Workshop on Patch-Based Techniques in Medical Imaging], Lecture Notes in Computer Science 10530, 20–28, Springer; (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [7].Shao M, Carass A, Li X, Dewey BE, Blitz AM, Prince JL, and Ellingsen LM, “Multi-atlas segmentation of the hydrocephalus brain using an adaptive ventricle atlas,” in [Proceedings of SPIE Medical Imaging (SPIE-MI 2018), Houstan, TX, February 10 – 15, 2018], 10578, 105780F–105780F–7 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [8].Shao M, Han S, Carass A, Li X, Dewey BE, Blitz AM, Prince JL, and Ellingsen LM, “Shortcomings of Ventricle Segmentation Using Deep Convolutional Networks,” in [Understanding and Interpreting Machine Learning in Medical Image Computing Applications. MLCN 2018, DLF 2018, IMIMIC 2018], Lecture Notes in Computer Science 11038, 79–86, Springer; Berlin Heidelberg: (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [9].Shao M, Han S, Carass A, Li X, Blitz AM, Shin J, Prince JL, and Ellingsen LM, “Brain ventricle parcellation using a deep neural network: Application to patients with ventriculomegaly,” NeuroImage: Clinical 23, 101871 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [10].Atlason HE, Shao M, Robertsson V, Sigurdsson S, Gudnason V, Prince JL, and Ellingsen LM, “Large-scale parcellation of the ventricular system using convolutional neural networks,” in [Proceedings of SPIE Medical Imaging (SPIE-MI 2019), San Diego, CA, February 16 – 21, 2019], 10953, 111–117 (2019). [Google Scholar]
  • [11].Atlason HE, Love A, Robertsson V, Blitz AM, Sigurdsson S, Gudnason V, and Ellingsen LM, “A joint ventricle and WMH segmentation from MRI for evaluation of healthy and pathological changes in the aging brain,” PLOS 17(9), e0274212 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [12].Ren X, Huo J, Xuan K, Wei D, Zhang L, and Wang Q, “Robust brain magnetic resonance image segmentation for hydrocephalus patients: Hard and soft attention,” in [17th International Symposium on Biomedical Imaging (ISBI 2020)], 385–389 (2020). [Google Scholar]
  • [13].Fischl B, “Freesurfer,” Neuroimage 62(2), 774–781 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [14].Dubost F, de Bruijne M, Nardin M, Dalca AV, Donahue KL, Giese A-K, Etherton MR, Wu O, de Groot M, Niessen W, Vernooij M, Rost NS, and Schirmer MD, “Multi-atlas image registration of clinical data with automated quality assessment using ventricle segmentation,” Medical Image Analysis 63, 101698 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [15].Ye F, Wang Z, Hu K, Zhu S, and Gao X, “Automated Segmentation of Lateral Ventricle in MR Images Using Multi-scale Feature Fusion Convolutional Neural Network,” in [International Conference on Pattern Recognition], 330–338 (2021). [Google Scholar]
  • [16].van Doormaal JAM, Fick T, Ali M, Köllen M, van der Kuijp V, and van Doormaal TPC, “Fully automatic adaptive meshing based segmentation of the ventricular system for augmented reality visualization and navigation,” World Neurosurgery 156, e9–e24 (2021). [DOI] [PubMed] [Google Scholar]
  • [17].Li Z, Li Z, Chen Q, Ramos A, Zhang J, Boudreaux JP, Thiagarajan R, Bren-Mattison Y, Dunham ME, McWhorter AJ, Li X, Feng J-M, Li Y, Yao S, and Xu J, “Detection of pancreatic cancer by convolutional-neural-network-assisted spontaneous raman spectroscopy with critical feature visualization,” Neural Networks 144, 455–464 (2021). [DOI] [PubMed] [Google Scholar]
  • [18].Li Z, Li Z, Yao L, Chen Q, Zhang J, Li X, Feng J-M, Li Y, and Xu J, “Multiple-Inputs Convolutional Neural Network for COVID-19 Classification and Critical Region Screening From Chest X-ray Radiographs: Model Development and Performance Evaluation,” JMIR Bioinformatics and Biotechnology 3(1), e36660 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [19].Wang Y and Yi J, “Deep learning based image registration method: with application to Scanning Laser Ophthalmoscopy (SLO) longitudinal images,” in [Proceedings of SPIE Medical Imaging (SPIE-MI 2023), San Diego, CA, February 19 – 23, 2023], (2023). [Google Scholar]
  • [20].Zhou P, Liu Z, Wu H, Wang Y, Lei Y, and Abbaszadeh S, “Automatically detecting bregma and lambda points in rodent skull anatomy images,” PloS one 15(12), e0244378 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [21].Wang Y, Feng A, Xue Y, Zuo L, Liu Y, Blitz AM, Luciano MG, Carass A, and Prince JL, “Automated Ventricle Parcellation and Evan’s Ratio Computation in Pre- and Post-Surgical Ventriculomegaly,” in [ISBI23], (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [22].Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, and Gee JC, “N4ITK: Improved N3 Bias Correction,” IEEE Trans. Med. Imag 29(6), 1310–1320 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [23].Iglesias JE, Liu C, Thompson PM, and Tu Z, “Robust brain extraction across datasets and comparison with publicly available methods,” TMI 30(9), 1617–1634 (2011). [DOI] [PubMed] [Google Scholar]
  • [24].Reinhold JC, Dewey BE, Carass A, and Prince JL, “Evaluating the Impact of Intensity Normalization on MR Image Synthesis,” in [Proceedings of SPIE Medical Imaging (SPIE-MI 2019), San Diego, CA, February 16 – 21, 2019], 10949, 109493H (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES