The Choroid Plexus (ChP) is a vascular tissue located inside the brain ventricles. The increased interest on the ChP is related to the recent discoveries on its immunological function in inflammatory processes. ChP volume (ChPV) measured by T1-w MRI has been observed altered in several neurological disorders (i.e., Multiple Sclerosis, Major Depressive Disorder, Alzheimer’s Disease, psychotic disorders). Therefore, ChPV can become a promising biomarker to improve the understanding of neurological diseases. However, the manual segmentation of ChP, that is the ground truth (GT), is time-consuming and affected by inter-operator variability due to the complexity of the task. FreeSurfer (FS) and the Gaussian Mixture Model Method (GMM) are the automatic methods proposed in literature. The aim of this work is to propose a method for the completely automatic, accurate, reliable, and fast semantic segmentation of the ChP based on novel deep learning neural networks (DNN). The main goal is to find the combinations of parameters that maximize the performance indices with respect to the gold-standard manual segmentation depicted over T1-w MRI image with contrast injection, without the use of sequences with contrast agents to make this task less invasive for the patient. The dataset analyzed is composed of 60 relapsing-remitting Multiple Sclerosis (RR-MS) patients, divided into a training set (45) and a validation set (15). The tested DNN are: 3D U-Net, V-Net, nnU-Net and UNETR. The training parameters and configurations that have been tested are: the input MRI sequence (T1-w, FLAIR, FLAIR+T1-w) and the GT segmentation (respectively, T1-w, FLAIR, T1-w or FLAIR; gold-standard cT1-w), the preprocessing with data augmentation transformations, the patch size (64x64x64, 96x96x96, 128x128x128) and the loss function (Cross-Entropy, Weighted Cross-Entropy, Dice, Dice-CE). The analyzed performance indices are: Dice Coefficient, Jaccard Index, 95% Hausdorff Distance, Percentage Volume Difference, Root Mean Squared Error (RMSE). The preliminary analysis over the GT of T1-w and FLAIR sequences, and the two automatic segmentations of FS and GMM with respect to the gold-standard one has demonstrated that FS and GMM are quite inaccurate with respect to the manual segmentations. This consideration has corroborated the need to propose an alternative automatic method to segment the ChP. Moreover, on one hand, T1-w sequence is to be preferred to use the ChPV as a quantitative biomarker, because it has the lower Percentage Volume Difference; on the other hand, FLAIR sequence lowers the variability of the resulted segmentation, as showed by the higher Dice Coefficient. The training analyses over all 672 possible combinations of DNNs have shown the better performances of nnU-Net and UNETR during the segmentation task. It was not possible to delineate the best combination of DNN parameters that could be equally suitable for each performance indices. Nevertheless, the most significant observations are that it is suggested avoiding the use of V-Net and Weighted Cross-Entropy. Making the comparison with the gold-standard segmentation, UNETR is slightly superior to nnU-Net and has brought to a Percentage Volume Difference on the validation set around 8% over T1-w images, trained both with cT1-w MSeg and T1-wMSeg. These results are remarkable and let its use on large clinical dataset, where the magnitude of Volume Difference between MS patients and healthy controls is around 21%. To conclude, UNETR is a reliable tool for the segmentation of the ChP using the T1-w images and shows its promising usefulness to establish a new neuroimaging biomarker without the use of invasive techniques.

The Choroid Plexus (ChP) is a vascular tissue located inside the brain ventricles. The increased interest on the ChP is related to the recent discoveries on its immunological function in inflammatory processes. ChP volume (ChPV) measured by T1-w MRI has been observed altered in several neurological disorders (i.e., Multiple Sclerosis, Major Depressive Disorder, Alzheimer’s Disease, psychotic disorders). Therefore, ChPV can become a promising biomarker to improve the understanding of neurological diseases. However, the manual segmentation of ChP, that is the ground truth (GT), is time-consuming and affected by inter-operator variability due to the complexity of the task. FreeSurfer (FS) and the Gaussian Mixture Model Method (GMM) are the automatic methods proposed in literature. The aim of this work is to propose a method for the completely automatic, accurate, reliable, and fast semantic segmentation of the ChP based on novel deep learning neural networks (DNN). The main goal is to find the combinations of parameters that maximize the performance indices with respect to the gold-standard manual segmentation depicted over T1-w MRI image with contrast injection, without the use of sequences with contrast agents to make this task less invasive for the patient. The dataset analyzed is composed of 60 relapsing-remitting Multiple Sclerosis (RR-MS) patients, divided into a training set (45) and a validation set (15). The tested DNN are: 3D U-Net, V-Net, nnU-Net and UNETR. The training parameters and configurations that have been tested are: the input MRI sequence (T1-w, FLAIR, FLAIR+T1-w) and the GT segmentation (respectively, T1-w, FLAIR, T1-w or FLAIR; gold-standard cT1-w), the preprocessing with data augmentation transformations, the patch size (64x64x64, 96x96x96, 128x128x128) and the loss function (Cross-Entropy, Weighted Cross-Entropy, Dice, Dice-CE). The analyzed performance indices are: Dice Coefficient, Jaccard Index, 95% Hausdorff Distance, Percentage Volume Difference, Root Mean Squared Error (RMSE). The preliminary analysis over the GT of T1-w and FLAIR sequences, and the two automatic segmentations of FS and GMM with respect to the gold-standard one has demonstrated that FS and GMM are quite inaccurate with respect to the manual segmentations. This consideration has corroborated the need to propose an alternative automatic method to segment the ChP. Moreover, on one hand, T1-w sequence is to be preferred to use the ChPV as a quantitative biomarker, because it has the lower Percentage Volume Difference; on the other hand, FLAIR sequence lowers the variability of the resulted segmentation, as showed by the higher Dice Coefficient. The training analyses over all 672 possible combinations of DNNs have shown the better performances of nnU-Net and UNETR during the segmentation task. It was not possible to delineate the best combination of DNN parameters that could be equally suitable for each performance indices. Nevertheless, the most significant observations are that it is suggested avoiding the use of V-Net and Weighted Cross-Entropy. Making the comparison with the gold-standard segmentation, UNETR is slightly superior to nnU-Net and has brought to a Percentage Volume Difference on the validation set around 8% over T1-w images, trained both with cT1-w MSeg and T1-wMSeg. These results are remarkable and let its use on large clinical dataset, where the magnitude of Volume Difference between MS patients and healthy controls is around 21%. To conclude, UNETR is a reliable tool for the segmentation of the ChP using the T1-w images and shows its promising usefulness to establish a new neuroimaging biomarker without the use of invasive techniques.

Confronto tra architetture deep learning per la segmentazione semantica del plesso coroideo a partire da immagini di risonanza magnetica cerebrale: applicazione alla sclerosi multipla

VISANI, VALENTINA
2021/2022

Abstract

The Choroid Plexus (ChP) is a vascular tissue located inside the brain ventricles. The increased interest on the ChP is related to the recent discoveries on its immunological function in inflammatory processes. ChP volume (ChPV) measured by T1-w MRI has been observed altered in several neurological disorders (i.e., Multiple Sclerosis, Major Depressive Disorder, Alzheimer’s Disease, psychotic disorders). Therefore, ChPV can become a promising biomarker to improve the understanding of neurological diseases. However, the manual segmentation of ChP, that is the ground truth (GT), is time-consuming and affected by inter-operator variability due to the complexity of the task. FreeSurfer (FS) and the Gaussian Mixture Model Method (GMM) are the automatic methods proposed in literature. The aim of this work is to propose a method for the completely automatic, accurate, reliable, and fast semantic segmentation of the ChP based on novel deep learning neural networks (DNN). The main goal is to find the combinations of parameters that maximize the performance indices with respect to the gold-standard manual segmentation depicted over T1-w MRI image with contrast injection, without the use of sequences with contrast agents to make this task less invasive for the patient. The dataset analyzed is composed of 60 relapsing-remitting Multiple Sclerosis (RR-MS) patients, divided into a training set (45) and a validation set (15). The tested DNN are: 3D U-Net, V-Net, nnU-Net and UNETR. The training parameters and configurations that have been tested are: the input MRI sequence (T1-w, FLAIR, FLAIR+T1-w) and the GT segmentation (respectively, T1-w, FLAIR, T1-w or FLAIR; gold-standard cT1-w), the preprocessing with data augmentation transformations, the patch size (64x64x64, 96x96x96, 128x128x128) and the loss function (Cross-Entropy, Weighted Cross-Entropy, Dice, Dice-CE). The analyzed performance indices are: Dice Coefficient, Jaccard Index, 95% Hausdorff Distance, Percentage Volume Difference, Root Mean Squared Error (RMSE). The preliminary analysis over the GT of T1-w and FLAIR sequences, and the two automatic segmentations of FS and GMM with respect to the gold-standard one has demonstrated that FS and GMM are quite inaccurate with respect to the manual segmentations. This consideration has corroborated the need to propose an alternative automatic method to segment the ChP. Moreover, on one hand, T1-w sequence is to be preferred to use the ChPV as a quantitative biomarker, because it has the lower Percentage Volume Difference; on the other hand, FLAIR sequence lowers the variability of the resulted segmentation, as showed by the higher Dice Coefficient. The training analyses over all 672 possible combinations of DNNs have shown the better performances of nnU-Net and UNETR during the segmentation task. It was not possible to delineate the best combination of DNN parameters that could be equally suitable for each performance indices. Nevertheless, the most significant observations are that it is suggested avoiding the use of V-Net and Weighted Cross-Entropy. Making the comparison with the gold-standard segmentation, UNETR is slightly superior to nnU-Net and has brought to a Percentage Volume Difference on the validation set around 8% over T1-w images, trained both with cT1-w MSeg and T1-wMSeg. These results are remarkable and let its use on large clinical dataset, where the magnitude of Volume Difference between MS patients and healthy controls is around 21%. To conclude, UNETR is a reliable tool for the segmentation of the ChP using the T1-w images and shows its promising usefulness to establish a new neuroimaging biomarker without the use of invasive techniques.
2021
A comparison of deep learning architectures for the semantic segmentation of choroid plexus from brain MRI: application to multiple sclerosis
The Choroid Plexus (ChP) is a vascular tissue located inside the brain ventricles. The increased interest on the ChP is related to the recent discoveries on its immunological function in inflammatory processes. ChP volume (ChPV) measured by T1-w MRI has been observed altered in several neurological disorders (i.e., Multiple Sclerosis, Major Depressive Disorder, Alzheimer’s Disease, psychotic disorders). Therefore, ChPV can become a promising biomarker to improve the understanding of neurological diseases. However, the manual segmentation of ChP, that is the ground truth (GT), is time-consuming and affected by inter-operator variability due to the complexity of the task. FreeSurfer (FS) and the Gaussian Mixture Model Method (GMM) are the automatic methods proposed in literature. The aim of this work is to propose a method for the completely automatic, accurate, reliable, and fast semantic segmentation of the ChP based on novel deep learning neural networks (DNN). The main goal is to find the combinations of parameters that maximize the performance indices with respect to the gold-standard manual segmentation depicted over T1-w MRI image with contrast injection, without the use of sequences with contrast agents to make this task less invasive for the patient. The dataset analyzed is composed of 60 relapsing-remitting Multiple Sclerosis (RR-MS) patients, divided into a training set (45) and a validation set (15). The tested DNN are: 3D U-Net, V-Net, nnU-Net and UNETR. The training parameters and configurations that have been tested are: the input MRI sequence (T1-w, FLAIR, FLAIR+T1-w) and the GT segmentation (respectively, T1-w, FLAIR, T1-w or FLAIR; gold-standard cT1-w), the preprocessing with data augmentation transformations, the patch size (64x64x64, 96x96x96, 128x128x128) and the loss function (Cross-Entropy, Weighted Cross-Entropy, Dice, Dice-CE). The analyzed performance indices are: Dice Coefficient, Jaccard Index, 95% Hausdorff Distance, Percentage Volume Difference, Root Mean Squared Error (RMSE). The preliminary analysis over the GT of T1-w and FLAIR sequences, and the two automatic segmentations of FS and GMM with respect to the gold-standard one has demonstrated that FS and GMM are quite inaccurate with respect to the manual segmentations. This consideration has corroborated the need to propose an alternative automatic method to segment the ChP. Moreover, on one hand, T1-w sequence is to be preferred to use the ChPV as a quantitative biomarker, because it has the lower Percentage Volume Difference; on the other hand, FLAIR sequence lowers the variability of the resulted segmentation, as showed by the higher Dice Coefficient. The training analyses over all 672 possible combinations of DNNs have shown the better performances of nnU-Net and UNETR during the segmentation task. It was not possible to delineate the best combination of DNN parameters that could be equally suitable for each performance indices. Nevertheless, the most significant observations are that it is suggested avoiding the use of V-Net and Weighted Cross-Entropy. Making the comparison with the gold-standard segmentation, UNETR is slightly superior to nnU-Net and has brought to a Percentage Volume Difference on the validation set around 8% over T1-w images, trained both with cT1-w MSeg and T1-wMSeg. These results are remarkable and let its use on large clinical dataset, where the magnitude of Volume Difference between MS patients and healthy controls is around 21%. To conclude, UNETR is a reliable tool for the segmentation of the ChP using the T1-w images and shows its promising usefulness to establish a new neuroimaging biomarker without the use of invasive techniques.
deep learning
choroid plexus
multiple sclerosis
MRI
segmentation
File in questo prodotto:
File Dimensione Formato  
Visani_Valentina.pdf

Open Access dal 15/04/2023

Dimensione 9.36 MB
Formato Adobe PDF
9.36 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/29250