Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9434075
C. Meyer, V. Mallouh, D. Spehner, É. Baudrier, P. Schultz, B. Naegel
Focused Ion Beam milling combined with Scanning Electron Microscopy (FIB-SEM) technique is an electron microscopy imaging method that offers the possibility of acquiring 3D isotropic images of biological structures at the nanometric scale. Automated image segmentation is required for morphological analysis of huge image stacks and to save time consuming manual intervention. Current methods are either specific to data and organelles or lack accuracy. We propose a robust multi-class semantic segmentation method for FIBSEM images, based on deep neural networks. We evaluate and compare our proposed method on two FIB-SEM images, for the segmentation of mitochondria, cell membrane and endoplasmic reticulum. We achieve results close to inter-expert variability.
{"title":"Automatic Multi Class Organelle Segmentation For Cellular Fib-Sem Images","authors":"C. Meyer, V. Mallouh, D. Spehner, É. Baudrier, P. Schultz, B. Naegel","doi":"10.1109/ISBI48211.2021.9434075","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9434075","url":null,"abstract":"Focused Ion Beam milling combined with Scanning Electron Microscopy (FIB-SEM) technique is an electron microscopy imaging method that offers the possibility of acquiring 3D isotropic images of biological structures at the nanometric scale. Automated image segmentation is required for morphological analysis of huge image stacks and to save time consuming manual intervention. Current methods are either specific to data and organelles or lack accuracy. We propose a robust multi-class semantic segmentation method for FIBSEM images, based on deep neural networks. We evaluate and compare our proposed method on two FIB-SEM images, for the segmentation of mitochondria, cell membrane and endoplasmic reticulum. We achieve results close to inter-expert variability.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"8 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125334966","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9433916
Yu Luo, Haoyang Chen, Jicong Zhang
The intensive practice of specific cognitive activities can lead to improvements of relevant cognitive capability in human beings, which may transfer to gain in untrained activities. Although there are a growing number of studies investigating the behavioral benefits of attention training in mind wandering, few studies have directly examined the neurophysiological basis of the training effects. Here using 128-channel electroencephalography (EEG), we examined whether the tactile training can reduce the mind wandering as measured by the sustained attention to response task (SART), and how the dynamic neurophysiological connectivity changes following training in young adults. The trainees showed significantly less occurrence of mind wandering after the five-day tactile training. Furthermore, the functional connectivity within and between the frontal and parietal regions was enhanced after training. Our findings suggest that the tactile training-induced brain plasticity may provide new therapeutic strategies for attention-related disorders.
{"title":"Enhanced Connectivity and Reduced Mind Wandering after Tactile Training in Young Adults","authors":"Yu Luo, Haoyang Chen, Jicong Zhang","doi":"10.1109/ISBI48211.2021.9433916","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9433916","url":null,"abstract":"The intensive practice of specific cognitive activities can lead to improvements of relevant cognitive capability in human beings, which may transfer to gain in untrained activities. Although there are a growing number of studies investigating the behavioral benefits of attention training in mind wandering, few studies have directly examined the neurophysiological basis of the training effects. Here using 128-channel electroencephalography (EEG), we examined whether the tactile training can reduce the mind wandering as measured by the sustained attention to response task (SART), and how the dynamic neurophysiological connectivity changes following training in young adults. The trainees showed significantly less occurrence of mind wandering after the five-day tactile training. Furthermore, the functional connectivity within and between the frontal and parietal regions was enhanced after training. Our findings suggest that the tactile training-induced brain plasticity may provide new therapeutic strategies for attention-related disorders.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130460331","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9434090
Giammarco La Barbera, P. Gori, Haithem Boussaid, Bruno Belucci, A. Delmonte, Jeanne Goulin, S. Sarnacki, L. Rouet, I. Bloch
Due to a high heterogeneity in pose and size and to a limited number of available data, segmentation of pediatric images is challenging for deep learning methods. In this work, we propose a new CNN architecture that is pose and scale invariant thanks to the use of Spatial Transformer Network (STN). Our architecture is composed of three sequential modules that are estimated together during training: (i) a regression module to estimate a similarity matrix to normalize the input image to a reference one; (ii) a differentiable module to find the region of interest to segment; (iii) a segmentation module, based on the popular UNet architecture, to delineate the object. Unlike the original UNet, which strives to learn a complex mapping, including pose and scale variations, from a finite training dataset, our segmentation module learns a simpler mapping focusing on images with normalized pose and size. Furthermore, the use of an automatic bounding box detection through STN allows saving time and especially memory, while keeping similar performance. We test the proposed method in kidney and renal tumor segmentation on abdominal pediatric CT scanners. Results indicate that the estimated STN homogenization of size and pose accelerates the segmentation (25h), compared to standard data-augmentation (33h), while obtaining a similar quality for the kidney (88.01% of Dice score) and improving the renal tumor delineation (from 85.52% to 87.12%).
{"title":"Automatic Size And Pose Homogenization With Spatial Transformer Network To Improve And Accelerate Pediatric Segmentation","authors":"Giammarco La Barbera, P. Gori, Haithem Boussaid, Bruno Belucci, A. Delmonte, Jeanne Goulin, S. Sarnacki, L. Rouet, I. Bloch","doi":"10.1109/ISBI48211.2021.9434090","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9434090","url":null,"abstract":"Due to a high heterogeneity in pose and size and to a limited number of available data, segmentation of pediatric images is challenging for deep learning methods. In this work, we propose a new CNN architecture that is pose and scale invariant thanks to the use of Spatial Transformer Network (STN). Our architecture is composed of three sequential modules that are estimated together during training: (i) a regression module to estimate a similarity matrix to normalize the input image to a reference one; (ii) a differentiable module to find the region of interest to segment; (iii) a segmentation module, based on the popular UNet architecture, to delineate the object. Unlike the original UNet, which strives to learn a complex mapping, including pose and scale variations, from a finite training dataset, our segmentation module learns a simpler mapping focusing on images with normalized pose and size. Furthermore, the use of an automatic bounding box detection through STN allows saving time and especially memory, while keeping similar performance. We test the proposed method in kidney and renal tumor segmentation on abdominal pediatric CT scanners. Results indicate that the estimated STN homogenization of size and pose accelerates the segmentation (25h), compared to standard data-augmentation (33h), while obtaining a similar quality for the kidney (88.01% of Dice score) and improving the renal tumor delineation (from 85.52% to 87.12%).","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130349892","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Intrahepatic vascular separation on contrast-enhanced Magnetic Resonance (MR) images is indispensable for the hepatic tumor surgery. This paper presents an unsupervised frame-work based on structural saliency for automatically separating portal vein (PV) and hepatic vein (HV) from contrast-enhanced multi-phase MR images. In our work, we propose a new multi-scale filter based on statistics and shape information in the region of interest, called SSIROI, with which the vascular connectivity and saliency in the 3D hepatic region can be guaranteed. Experiments are conducted on clinical contrast-enhanced MR images, and the results show that our method achieves effective separation of intrahepatic vasculature by extracting the PV and HV from multi-phase images, and our proposed SSIROI filter outperforms state-of-the-art methods.
{"title":"A Structural Saliency-Based Approach for Automatic Intrahepatic Vascular Separation From Contrast-Enhanced Multi-Phase MR Images","authors":"Q. Guo, Hong Song, Jingfan Fan, Danni Ai, Jian Yang, Yuanjin Gao","doi":"10.1109/ISBI48211.2021.9433995","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9433995","url":null,"abstract":"Intrahepatic vascular separation on contrast-enhanced Magnetic Resonance (MR) images is indispensable for the hepatic tumor surgery. This paper presents an unsupervised frame-work based on structural saliency for automatically separating portal vein (PV) and hepatic vein (HV) from contrast-enhanced multi-phase MR images. In our work, we propose a new multi-scale filter based on statistics and shape information in the region of interest, called SSIROI, with which the vascular connectivity and saliency in the 3D hepatic region can be guaranteed. Experiments are conducted on clinical contrast-enhanced MR images, and the results show that our method achieves effective separation of intrahepatic vasculature by extracting the PV and HV from multi-phase images, and our proposed SSIROI filter outperforms state-of-the-art methods.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"35 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128451760","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9434159
Tran Minh Quan, Huynh Minh Thanh, Ta Duc Huy, Nguyen Do Trung Chanh, Nguyen Thi Hong Anh, Phan Hoan Vu, N. H. Nam, Tran Quy Tuong, Vu Minh Dien, B. Giang, Bui Huu Trung, S. Q. Truong
This work aims to fight against the current outbreak pandemic by developing a method to classify suspected infected COVID-19 cases. Driven by the urgency, due to the vastly increased number of patients and deaths worldwide, we rely on situationally pragmatic chest X-ray scans and state-of-the-art deep learning techniques to build a robust diagnosis for massive screening, early detection, and in-time isolation decision making. The proposed solution, X-ray Projected Generative Adversarial Network (XPGAN), addresses the most fundamental issue in training such a deep neural network on limited human-annotated datasets. By leveraging the generative adversarial network, we can synthesize a large amount of chest X-ray images with prior categories from more accurate 3D Computed Tomography data, including COVID-19, and jointly train a model with a few hundreds of positive samples. As a result, XPGAN outperforms the vanilla DenseNet121 models and other competing baselines trained on the same frontal chest X-ray images.
{"title":"XPGAN: X-Ray Projected Generative Adversarial Network For Improving Covid-19 Image Classification","authors":"Tran Minh Quan, Huynh Minh Thanh, Ta Duc Huy, Nguyen Do Trung Chanh, Nguyen Thi Hong Anh, Phan Hoan Vu, N. H. Nam, Tran Quy Tuong, Vu Minh Dien, B. Giang, Bui Huu Trung, S. Q. Truong","doi":"10.1109/ISBI48211.2021.9434159","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9434159","url":null,"abstract":"This work aims to fight against the current outbreak pandemic by developing a method to classify suspected infected COVID-19 cases. Driven by the urgency, due to the vastly increased number of patients and deaths worldwide, we rely on situationally pragmatic chest X-ray scans and state-of-the-art deep learning techniques to build a robust diagnosis for massive screening, early detection, and in-time isolation decision making. The proposed solution, X-ray Projected Generative Adversarial Network (XPGAN), addresses the most fundamental issue in training such a deep neural network on limited human-annotated datasets. By leveraging the generative adversarial network, we can synthesize a large amount of chest X-ray images with prior categories from more accurate 3D Computed Tomography data, including COVID-19, and jointly train a model with a few hundreds of positive samples. As a result, XPGAN outperforms the vanilla DenseNet121 models and other competing baselines trained on the same frontal chest X-ray images.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"39 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126088813","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9434136
Zixun Huang, Rui Zhao, Frank H. F. Leung, K. Lam, S. Ling, Juan Lyu, Sunetra Banerjee, T. Lee, De Yang, Y. Zheng
Ultrasound volume projection imaging (VPI) has shown to be appealing from a clinical perspective, because of its harmlessness, flexibility, and efficiency in scoliosis assessment. However, the limitations in hardware devices degrade the resultant image content with strong structured noise. Owing to the unavailability of reference data and the unpredictable degradation model, VPI image recovery is a challenging problem. In this paper, we propose a novel framework to learn the structured noise removal from unpaired samples. We introduce the attention mechanism into the generative adversarial network to enhance the learning by focusing on the salient corrupted patterns. We also present a dual adversarial learning strategy and integrate the denoiser with a segmentation model to produce the task-oriented noiseless estimation. Experimental results show that the proposed method can improve both the visual quality and the segmentation accuracy on spine images.
{"title":"DA-GAN: Learning Structured Noise Removal In Ultrasound Volume Projection Imaging For Enhanced Spine Segmentation","authors":"Zixun Huang, Rui Zhao, Frank H. F. Leung, K. Lam, S. Ling, Juan Lyu, Sunetra Banerjee, T. Lee, De Yang, Y. Zheng","doi":"10.1109/ISBI48211.2021.9434136","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9434136","url":null,"abstract":"Ultrasound volume projection imaging (VPI) has shown to be appealing from a clinical perspective, because of its harmlessness, flexibility, and efficiency in scoliosis assessment. However, the limitations in hardware devices degrade the resultant image content with strong structured noise. Owing to the unavailability of reference data and the unpredictable degradation model, VPI image recovery is a challenging problem. In this paper, we propose a novel framework to learn the structured noise removal from unpaired samples. We introduce the attention mechanism into the generative adversarial network to enhance the learning by focusing on the salient corrupted patterns. We also present a dual adversarial learning strategy and integrate the denoiser with a segmentation model to produce the task-oriented noiseless estimation. Experimental results show that the proposed method can improve both the visual quality and the segmentation accuracy on spine images.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"125 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127207869","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9434093
C. Román, N. López-López, J. Houenou, C. Poupon, J. F. Mangin, C. Hernández, P. Guevara
The study of the superficial white matter and its description is essential for the understanding of human brain function and the study of pathogenesis. However, the study of these fibers is still an incomplete task due to the high inter-subject variability and the size of this kind of fibers. In this work, a superficial white matter bundle identification based on fiber clustering was performed using probabilistic tractography on 100 subjects from the The Human Connectome Project (HCP) data, aligned with a non-linear registration. The method starts with an intra-subject clustering, followed by a segmentation of fibers connecting the precentral (PrC) and postcentral (PoC) regions, based on a ROI atlas. Due to the high amount of fibers, they were randomly separated into groups. An inter-subject clustering was applied on the fibers of each group, and then two clustering levels were applied to select the most reproducible bundles. Seven bundles per hemisphere were obtained, connecting the PrC and PoC regions. These were compared with bundles from previous atlases, showing in general more coverage and some bundles not found in previous atlases.
{"title":"Study Of Precentral-Postcentral Connections On Hcp Data Using Probabilistic Tractography And Fiber Clustering","authors":"C. Román, N. López-López, J. Houenou, C. Poupon, J. F. Mangin, C. Hernández, P. Guevara","doi":"10.1109/ISBI48211.2021.9434093","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9434093","url":null,"abstract":"The study of the superficial white matter and its description is essential for the understanding of human brain function and the study of pathogenesis. However, the study of these fibers is still an incomplete task due to the high inter-subject variability and the size of this kind of fibers. In this work, a superficial white matter bundle identification based on fiber clustering was performed using probabilistic tractography on 100 subjects from the The Human Connectome Project (HCP) data, aligned with a non-linear registration. The method starts with an intra-subject clustering, followed by a segmentation of fibers connecting the precentral (PrC) and postcentral (PoC) regions, based on a ROI atlas. Due to the high amount of fibers, they were randomly separated into groups. An inter-subject clustering was applied on the fibers of each group, and then two clustering levels were applied to select the most reproducible bundles. Seven bundles per hemisphere were obtained, connecting the PrC and PoC regions. These were compared with bundles from previous atlases, showing in general more coverage and some bundles not found in previous atlases.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"32 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126364950","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9433967
Minjeong Kim, Guorong Wu
Networks of biomarker covariance based on neuropathological events or neuro-degeneration degree is important to understand genetic influence and trophic reinforcement in the brain development/aging process. It is a common to quantiry the covariance of inter-subject biomarker profiles by linear correlation metrics such as Pearson’s correlation. Due to the heterogeneity and noise in the observed neurobiological data, however, it is difficult to construct a reliable covariance network using gross statistical measurement. To this, we propose a graph learning approach to infer the brain connectivity based on the harmonized inter-subject biomarker profiles. Specifically, we progressively estimate brain network until region-to-region connectivities reach the largest consensus of biomarker covariance across individuals. A better understanding of the network topology allows us to harmonize the neurobiological data effectively which eventually facilitates the graph inference. Since the network of biomarker covariance represents the region-wise associations in the entire population, we further promote diversity by adaptively penalizing the predominant influence from a group of biomarker profiles exhibiting statistically correlated patterns. We applied our method to the cortical thickness from MRI and amyloid-beta burden from PET images, which are biomarkers in Alzheimer’s disease (AD). Enhanced statistical power and replicability have been achieved by our approach in identifying network alterations between cognitive normal (CN) and AD cohorts.
{"title":"Constructing Reliable Network Of Biomarker Covariance By Joint Data Harmonization And Graph Learning","authors":"Minjeong Kim, Guorong Wu","doi":"10.1109/ISBI48211.2021.9433967","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9433967","url":null,"abstract":"Networks of biomarker covariance based on neuropathological events or neuro-degeneration degree is important to understand genetic influence and trophic reinforcement in the brain development/aging process. It is a common to quantiry the covariance of inter-subject biomarker profiles by linear correlation metrics such as Pearson’s correlation. Due to the heterogeneity and noise in the observed neurobiological data, however, it is difficult to construct a reliable covariance network using gross statistical measurement. To this, we propose a graph learning approach to infer the brain connectivity based on the harmonized inter-subject biomarker profiles. Specifically, we progressively estimate brain network until region-to-region connectivities reach the largest consensus of biomarker covariance across individuals. A better understanding of the network topology allows us to harmonize the neurobiological data effectively which eventually facilitates the graph inference. Since the network of biomarker covariance represents the region-wise associations in the entire population, we further promote diversity by adaptively penalizing the predominant influence from a group of biomarker profiles exhibiting statistically correlated patterns. We applied our method to the cortical thickness from MRI and amyloid-beta burden from PET images, which are biomarkers in Alzheimer’s disease (AD). Enhanced statistical power and replicability have been achieved by our approach in identifying network alterations between cognitive normal (CN) and AD cohorts.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"12 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127081909","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9433914
A. Adishesha, D. Vanselow, P. L. Rivière, Xiaolei Huang, K. Cheng
X-ray “Histotomography” built on the basic principles of CT can be used to create 3D images of zebrafish at resolutions one thousand times greater than CT, enabling the visualization of cell nuclei and other subcellular structures in 3D. Noise in the scans caused either through natural Xray phenomena or other distortions can lead to low accuracy in tasks related to detection and segmentation of anatomically significant objects. We evaluate the use of supervised Encoder-Decoder models for noise removal in projection and reconstruction domain images in absence of clean training targets. We propose the use of a Noise-2-Noise architecture with U-Net backbone along with structural similarity index loss as an addendum to help maintain and sharpen pathologically relevant details. We empirically show that our technique outperforms existing methods, with an average peak signal to noise ratio (PSNR) gain of 14. 50dB and 15. 05dB for noise removal in the reconstruction domain when trained without and with clean targets respectively. Using the same network architecture, we obtain a gain in structural similarity index (SSIM) in the projection domain by an average of 0.213 when trained without clean targets and 0.259 with clean targets. Additionally, by comparing reconstructions from denoised projections with those from original projections, we establish that noise removal in the projection domain is beneficial to improve the quality of reconstructed scans.
{"title":"Zebrafish Histotomography Noise Removal In Projection And Reconstruction Domains","authors":"A. Adishesha, D. Vanselow, P. L. Rivière, Xiaolei Huang, K. Cheng","doi":"10.1109/ISBI48211.2021.9433914","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9433914","url":null,"abstract":"X-ray “Histotomography” built on the basic principles of CT can be used to create 3D images of zebrafish at resolutions one thousand times greater than CT, enabling the visualization of cell nuclei and other subcellular structures in 3D. Noise in the scans caused either through natural Xray phenomena or other distortions can lead to low accuracy in tasks related to detection and segmentation of anatomically significant objects. We evaluate the use of supervised Encoder-Decoder models for noise removal in projection and reconstruction domain images in absence of clean training targets. We propose the use of a Noise-2-Noise architecture with U-Net backbone along with structural similarity index loss as an addendum to help maintain and sharpen pathologically relevant details. We empirically show that our technique outperforms existing methods, with an average peak signal to noise ratio (PSNR) gain of 14. 50dB and 15. 05dB for noise removal in the reconstruction domain when trained without and with clean targets respectively. Using the same network architecture, we obtain a gain in structural similarity index (SSIM) in the projection domain by an average of 0.213 when trained without clean targets and 0.259 with clean targets. Additionally, by comparing reconstructions from denoised projections with those from original projections, we establish that noise removal in the projection domain is beneficial to improve the quality of reconstructed scans.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"8 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114425193","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-04-13DOI: 10.1109/ISBI48211.2021.9433876
A. Chakravarty, Avik Kar, Ramanathan Sethuraman, D. Sheet
The shortage of Radiologists is inspiring the development of Deep Learning (DL) based solutions for detecting cardio, thoracic and pulmonary pathologies in Chest radiographs through multi-institutional collaborations. However, sharing the training data across multiple sites is often impossible due to privacy, ownership and technical challenges. Although Federated Learning (FL) has emerged as a solution to this, the large variations in disease prevalence and co-morbidity distributions across the sites may hinder proper training. We propose a DL architecture with a Convolutional Neural Network (CNN) followed by a Graph Neural Network (GNN) to address this issue. The CNN-GNN model is trained by modifying the Federated Averaging algorithm. The CNN weights are shared across all sites to extract robust features while separate GNN models are trained at each site to leverage the local co-morbidity dependencies for multi-label disease classification. The CheXpert dataset is partitioned across five sites to simulate the FL set up. Federated training did not show any significant drop in performance over centralized training. The site-specific GNN models also demonstrated their efficacy in modelling local disease co-occurrence statistics leading to an average area under the ROC curve of 0.79 with a 1.74% improvement.
{"title":"Federated Learning for Site Aware Chest Radiograph Screening","authors":"A. Chakravarty, Avik Kar, Ramanathan Sethuraman, D. Sheet","doi":"10.1109/ISBI48211.2021.9433876","DOIUrl":"https://doi.org/10.1109/ISBI48211.2021.9433876","url":null,"abstract":"The shortage of Radiologists is inspiring the development of Deep Learning (DL) based solutions for detecting cardio, thoracic and pulmonary pathologies in Chest radiographs through multi-institutional collaborations. However, sharing the training data across multiple sites is often impossible due to privacy, ownership and technical challenges. Although Federated Learning (FL) has emerged as a solution to this, the large variations in disease prevalence and co-morbidity distributions across the sites may hinder proper training. We propose a DL architecture with a Convolutional Neural Network (CNN) followed by a Graph Neural Network (GNN) to address this issue. The CNN-GNN model is trained by modifying the Federated Averaging algorithm. The CNN weights are shared across all sites to extract robust features while separate GNN models are trained at each site to leverage the local co-morbidity dependencies for multi-label disease classification. The CheXpert dataset is partitioned across five sites to simulate the FL set up. Federated training did not show any significant drop in performance over centralized training. The site-specific GNN models also demonstrated their efficacy in modelling local disease co-occurrence statistics leading to an average area under the ROC curve of 0.79 with a 1.74% improvement.","PeriodicalId":372939,"journal":{"name":"2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)","volume":"74 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-04-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121683412","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：