Article
Neurosciences
Yiqin Cao, Zhenyu Zhu, Yi Rao, Chenchen Qin, Di Lin, Qi Dou, Dong Ni, Yi Wang
Summary: EPReg is an edge-aware pyramidal deformable network for unsupervised volumetric registration. It utilizes multi-level feature pyramids and integrates edge information to enhance image structure alignment, enabling it to handle large deformations.
FRONTIERS IN NEUROSCIENCE
(2021)
Article
Computer Science, Artificial Intelligence
Brendan Kolisnik, Isaac Hogan, Farhana Zulkernine
Summary: We propose a hierarchical image classification model, Condition-CNN, which improves prediction accuracy and reduces training time by using the Teacher Forcing training algorithm and conditional probabilities. The validation results show that Condition-CNN achieves higher prediction accuracy for Level 1, 2, and 3 classes compared to other baseline CNN models.
EXPERT SYSTEMS WITH APPLICATIONS
(2021)
Article
Computer Science, Artificial Intelligence
Peng Wang, Yunqi Yan, Lijun Qian, Shiteng Suo, Jianrong Xu, Yi Guo, Yuanyuan Wang
Summary: Deep learning-based deformable image registration methods have shown great performance and fast run time. However, accurately estimating large topology-preserved deformation and utilizing important contextual information remains a challenge. In this study, we propose a novel unsupervised context-driven pyramid registration network, CPRNet, to address these issues by exploiting multi-scale spatial correlation and fusing deep contextual information. Our experiments on liver CT images and brain MR images demonstrate the effectiveness and accuracy of our proposed method in various datasets with fast run time. It outperforms existing learning-based registration methods while maintaining desirable topology preservation capability.
Article
Computer Science, Artificial Intelligence
Yunzhi Huang, Sahar Ahmad, Jingfan Fan, Dinggang Shen, Pew-Thian Yap
Summary: Deformable brain image registration aims to align anatomical structures in the presence of large and complex deformations. This study introduces a difficulty-aware model based on an attention mechanism to automatically identify challenging regions for better estimation of deformations. By incorporating the difficulty-aware model into a cascaded neural network, both global and local information are effectively leveraged for improved registration accuracy.
MEDICAL IMAGE ANALYSIS
(2021)
Article
Neurosciences
Michele Svanera, Sergio Benini, Dennis Bontempi, Lars Muckli
Summary: The article introduces an optimized neural network CEREBRUM-7T for automatic segmentation of 7T brain MRI, which can accurately and quickly segment brain MRI images for neuroimaging studies.
HUMAN BRAIN MAPPING
(2021)
Article
Computer Science, Artificial Intelligence
Khondaker Tasrif Noor, Antonio Robles-Kelly
Summary: In this paper, the authors propose H-CapsNet, a capsule network designed for hierarchical image classification. The network effectively captures hierarchical relationships using dedicated capsules for each class hierarchy. A modified hinge loss is utilized to enforce consistency among the involved hierarchies. Additionally, a strategy for dynamically adjusting training parameters is presented to achieve better balance between the class hierarchies. Experimental results demonstrate that H-CapsNet outperforms competing hierarchical classification networks.
PATTERN RECOGNITION
(2024)
Article
Computer Science, Information Systems
Hessam Sokooti, Sahar Yousefi, Mohamed S. Elmahdy, Boudewijn P. F. Lelieveldt, Marius Staring
Summary: In this paper, a supervised method using CNNs to predict registration misalignment is proposed. By leveraging hierarchical predictions and multi-resolution information through LSTM, the system achieves better registration results and higher accuracy in chest CT scans.
Article
Computer Science, Artificial Intelligence
Pablo Arratia Lopez, Hernan Mella, Sergio Uribe, Daniel E. Hurtado, Francisco Sahli Costabal
Summary: In this study, we propose a physics-informed neural network called WarpPINN for image registration to obtain local metrics of heart deformation. The network is informed about the near-incompressibility of cardiac tissue and computes cardiac strain using automatic differentiation. The algorithm outperforms current methodologies in landmark tracking, providing precise measurements of local cardiac deformations for the diagnosis of heart failure.
MEDICAL IMAGE ANALYSIS
(2023)
Article
Radiology, Nuclear Medicine & Medical Imaging
Zhenyu Zhu, Yiqin Cao, Chenchen Qin, Yi Rao, Di Lin, Qi Dou, Dong Ni, Yi Wang
Summary: The proposed network in this study combines affine and deformable registration methods for 3D medical image registration, achieving deformable registration in one forward pass without the need for pre-alignment. Experimental results show that the network outperformed state-of-the-art methods in terms of Dice index, Hausdorff distance, and average symmetric surface distance.
Article
Engineering, Biomedical
Gangcheng Cai, Huaying Liu, Wei Zou, Nan Hu, JiaJun Wang
Summary: In this paper, a deformable registration network (DR-Net) and a multi-scale cascading strategy are proposed for the registration of largely deformed 3D medical images. The DR-Net is constructed with a U-shaped convolutional neural network, a pyramidal input module, a light weighted sequential Inception module, and an SCAM convolutional attention module. The multi-scale cascading strategy integrates the deformation field information within and between sub-networks at different scales to synthesize the cascaded deformation fields.
BIOMEDICAL SIGNAL PROCESSING AND CONTROL
(2023)
Article
Engineering, Biomedical
Lijun Qian, Qing Zhou, Xiaohuan Cao, Wenjun Shen, Shiteng Suo, Shanshan Ma, Guoxiang Qu, Xuhua Gong, Yunqi Yan, Jianrong Xu, Luan Jiang
Summary: A cascade network framework composed of a de-enhancement network and a registration network is proposed for efficient registration of liver images in DCE-MRI. Experimental results show that the method improves efficiency while maintaining registration performance.
COMPUTERIZED MEDICAL IMAGING AND GRAPHICS
(2021)
Article
Computer Science, Interdisciplinary Applications
Lin Ge, Xingyue Wei, Yayu Hao, Jianwen Luo, Yan Xu
Summary: This study proposed an unsupervised structural feature guided convolutional neural network method for the registration of multiple stained images. Through the combination of low-resolution and high-resolution structural features, as well as a multi-scale strategy, it effectively overcame challenges such as repetitive texture and section missing.
IEEE TRANSACTIONS ON MEDICAL IMAGING
(2022)
Article
Biology
Xin Ma, Yajing Zhao, Yiping Lu, Peng Li, Xuanxuan Li, Nan Mei, Jiajun Wang, Daoying Geng, Lingxiao Zhao, Bo Yin
Summary: This paper proposes a deep learning model for accurate segmentation of meningiomas in magnetic resonance images. The model extracts features to achieve accurate tumor delineation, demonstrating high accuracy and generalizability.
COMPUTERS IN BIOLOGY AND MEDICINE
(2022)
Article
Neurosciences
Zilong Zeng, Tengda Zhao, Lianglong Sun, Yihe Zhang, Mingrui Xia, Xuhong Liao, Jiaying Zhang, Dinggang Shen, Li Wang, Yong He
Summary: In this study, a new 3D mixed-scale asymmetric convolutional segmentation network (3D-MASNet) was proposed for tissue segmentation of 6-month-old infant brain MRI images. Compared to traditional single-scale symmetric convolutions, this approach demonstrated better accuracy and achieved the best performance in the evaluation.
HUMAN BRAIN MAPPING
(2023)
Article
Computer Science, Hardware & Architecture
Ravi Shanker, Heet Sankesara, Surendra Nagar, Mahua Bhattacharya
Summary: In recent years, complex deep learning-based architectures have been developed for medical image registration, but their compatibility issues and high memory requirements limit their scalability. To address this problem, we propose a resource-efficient and structure-preserving network (RESPNet) for medical image registration. The architecture utilizes the structural properties of images for 2D and 3D registration and outperforms the current state-of-the-art methods in terms of memory requirements, dice score, and processing time.
JOURNAL OF SUPERCOMPUTING
(2023)
Article
Computer Science, Artificial Intelligence
Hong Song, Lei Chen, Yutao Cui, Qiang Li, Qi Wang, Jingfan Fan, Jian Yang, Le Zhang
Summary: In this paper, the authors propose a cascaded multi-supervision convolutional neural network (CMSNet) to remove low-dose perfusion noise in CT images and Rician noise in MR images. The network consists of a multi-supervision network (MSNet) and a refinement network, which are used for noise prediction and denoising operations. Experimental results demonstrate the promising performance of the proposed model under different noise levels.
Article
Engineering, Biomedical
Chan Wu, Tianyu Fu, Yifan Wang, Yucong Lin, Yan Wang, Danni Ai, Jingfan Fan, Hong Song, Jian Yang
Summary: This article proposes a fusion Siamese network with drift correction for target tracking in ultrasound sequences. By generating response maps using cross-correlation between convolution layers at different resolutions and combining a correction strategy, the method effectively corrects target drift. Experimental results demonstrate its high tracking accuracy on ultrasound image data.
PHYSICS IN MEDICINE AND BIOLOGY
(2022)
Article
Biochemical Research Methods
Shiyuan Liu, Jingfan Fan, Danni Ai, Hong Song, Tianyu Fu, Yongtian Wang, Jian Yang
Summary: Feature matching is crucial in obtaining the surface morphology of soft tissues in intraoperative endoscopy images. We proposed an adaptive gradient-preserving method to enhance the visual features of texture-less images. Experimental results demonstrate the effectiveness of our method in feature point extraction and surface reconstruction.
BIOMEDICAL OPTICS EXPRESS
(2022)
Editorial Material
Biophysics
Hui Zhang, Daniel C. Alexander, Dinggang Shen, Pew-Thian Yap
NMR IN BIOMEDICINE
(2022)
Article
Engineering, Biomedical
Wenjie Li, Jingfan Fan, Yating Li, Pengcheng Hao, Yucong Lin, Tianyu Fu, Danni Ai, Hong Song, Jian Yang
Summary: This study presents a physics model driven semi-supervised learning framework for high-quality pixel-wise endoscopic image enhancement in order to address common issues in endoscopic surgery, such as smoke, uneven lighting, and color deviation. The proposed network integrates specific physical imaging defect models with the CycleGAN framework and achieves superior performance compared to state-of-the-art methods. Transfer learning is also applied to overcome data scarcity in endoscope enhancement tasks, resulting in improved network performance.
PHYSICS IN MEDICINE AND BIOLOGY
(2022)
Article
Computer Science, Interdisciplinary Applications
Xu Chen, Tianshu Kuang, Hannah Deng, Steve H. Fung, Jaime Gateno, James J. Xia, Pew-Thian Yap
Summary: This paper introduces a dual-attention domain-adaptative segmentation network for cross-modality medical image segmentation, which shows superior performance in challenging tasks.
IEEE TRANSACTIONS ON MEDICAL IMAGING
(2022)
Article
Biochemical Research Methods
Sahar Ahmad, Ye Wu, Zhengwang Wu, Kim-Han Thung, Siyuan Liu, Weili Lin, Gang Li, Li Wang, Pew-Thian Yap
Summary: Brain atlases serve as spatial references for integrating, processing, and analyzing brain features from different individuals, sources, and scales. This study introduces a collection of joint surface-volume atlases that map the postnatal development of the human brain in a densely spatiotemporal manner from two weeks to two years old. These atlases provide normative patterns and capture key traits of early brain development, enabling the identification of deviations from normal developmental trajectories. By facilitating the mapping of diverse features of the infant brain to a common reference frame, these atlases enhance our understanding of early structural and functional development by enabling precise quantification of cortical and subcortical changes.
Article
Computer Science, Interdisciplinary Applications
Feng Cheng, Yilin Liu, Yong Chen, Pew-Thian Yap
Summary: Magnetic Resonance Fingerprinting (MRF) is a new quantitative imaging framework for simultaneous measurement of multiple tissue properties. A graph convolution network (GCN) is developed to replace GRAPPA for accurate tissue quantification in MRF. The GCN achieves up to 6-fold acceleration and enables the acquisition of whole-brain 3D MRF data with 1mm isotropic resolution in 3 minutes, or submillimeter 3D MRF data (0.8mm) in 5 minutes, greatly improving the feasibility of MRF in clinical settings.
IEEE TRANSACTIONS ON MEDICAL IMAGING
(2023)
Article
Computer Science, Artificial Intelligence
Geng Chen, Yoonmi Hong, Khoi Minh Huynh, Pew-Thian Yap
Summary: This study proposes two novel loss functions, microstructural loss and spherical variance loss, to explicitly consider the quality of both the predicted DMRI data and derived diffusion scalars. Applying these loss functions to multi-shell data prediction and angular resolution enhancement, evaluation results show that both microstructural loss and spherical variance loss improve the quality of derived diffusion scalars.
MEDICAL IMAGE ANALYSIS
(2023)
Article
Computer Science, Information Systems
Hao Guan, Ling Yue, Pew-Thian Yap, Shifu Xiao, Andrea Bozoki, Mingxia Liu
Summary: Subjective cognitive decline (SCD) is an early stage of Alzheimer's disease (AD) that occurs before mild cognitive impairment (MCI). It can progress to MCI and eventually to AD. Therefore, early identification of progressive SCD using neuroimaging techniques like structural MRI is crucial for early intervention of AD. This study proposes an interpretable autoencoder model with domain transfer learning (IADT) to predict the progression of SCD. The model leverages MRIs from both the target domain (SCD) and auxiliary domains (AD and NC), and uses an attention mechanism to automatically locate disease-related brain regions, achieving good interpretability. The IADT model is easy to train and test, requires only 5 to 10 seconds on CPUs, and is suitable for medical tasks with small datasets. Extensive experiments on publicly available and private datasets have demonstrated the effectiveness of the proposed method.
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
(2023)
Article
Computer Science, Artificial Intelligence
Zhanlong Yang, Linzhi Yang, Geng Chen, Pew-Thian Yap
Summary: In this paper, a novel local image descriptor called IPCET is proposed, which is based on the phase and amplitude information of PCET. The IPCET descriptor is robust to both photometric and geometric transformations, and outperforms cutting-edge moment-based descriptors according to extensive experiments.
PATTERN RECOGNITION
(2023)
Article
Computer Science, Artificial Intelligence
Yunzhi Huang, Sahar Ahmad, Luyi Han, Shuai Wang, Zhengwang Wu, Weili Lin, Gang Li, Li Wang, Pew-Thian Yap
Summary: This paper proposes a deep learning framework to predict missing scans in longitudinal infant studies. The model, called MGAN, uses image translation and quality-guided learning strategies to accurately predict tissue contrasts and anatomical details. Experimental results show that MGAN outperforms existing GANs in this task.
PATTERN RECOGNITION
(2023)
Article
Multidisciplinary Sciences
Xiaoyang Chen, Liangqiong Qu, Yifang Xie, Sahar Ahmad, Pew-Thian Yap
Summary: Brain MRI provides critical soft tissue contrasts for disease diagnosis and neuroscience research. Ultra-high field strength 7T MRI allows for higher resolution and better tissue contrast and SNR, but its high costs limit its adoption. To obtain higher-quality images without 7T MRI, algorithms for synthesizing 7T images from 3T images are being developed. A dataset of paired 3T and 7T MR images of 10 healthy subjects is made available to facilitate the development and evaluation of 3T-to-7T MR image synthesis models.
Article
Computer Science, Artificial Intelligence
Qing Guo, Hong Song, Jingfan Fan, Danni Ai, Yuanjin Gao, Xiaoling Yu, Jian Yang
Summary: In this paper, a unified framework is proposed for automatically and robustly segmenting 3D hepatic vein (HV) and portal vein (PV) from multi-phase MRI images. The framework considers the change and appearance caused by vascular flow events and improves the segmentation performance through clustering and decision-making methods. The evaluation results demonstrate that the proposed framework outperforms existing methods.
IEEE TRANSACTIONS ON IMAGE PROCESSING
(2022)
Article
Engineering, Biomedical
Long Shao, Tianyu Fu, Zhao Zheng, Zehua Zhao, Lele Ding, Jingfan Fan, Hong Song, Tao Zhang, Jian Yang
Summary: This paper proposes a method to track the position of the patient directly using video data from a single camera, which achieves noninvasive, real time, and high positioning accuracy in facial repair surgeries.
INTERNATIONAL JOURNAL OF COMPUTER ASSISTED RADIOLOGY AND SURGERY
(2022)
Article
Computer Science, Artificial Intelligence
Hong Liu, Dong Wei, Donghuan Lu, Xiaoying Tang, Liansheng Wang, Yefeng Zheng
Summary: This study proposes a framework based on hybrid 2D-3D convolutional neural networks for obtaining continuous 3D retinal layer surfaces from OCT volumes. The framework works well with both full and sparse annotations and utilizes alignment displacement vectors and layer segmentation to align the B-scans and segment the layers. Experimental results show that the framework outperforms state-of-the-art 2D deep learning methods in terms of layer segmentation accuracy and cross-B-scan 3D continuity.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Simon Oxenford, Ana Sofia Rios, Barbara Hollunder, Clemens Neudorfer, Alexandre Boutet, Gavin J. B. Elias, Jurgen Germann, Aaron Loh, Wissam Deeb, Bryan Salvato, Leonardo Almeida, Kelly D. Foote, Robert Amaral, Paul B. Rosenberg, David F. Tang-Wai, David A. Wolk, Anna D. Burke, Marwan N. Sabbagh, Stephen Salloway, M. Mallar Chakravarty, Gwenn S. Smith, Constantine G. Lyketsos, Michael S. Okun, William S., Zoltan Mari, Francisco A. Ponce, Andres Lozano, Wolf-Julian Neumann, Bassam Al-Fatly, Andreas Horn
Summary: Spatial normalization is a method to map subject brain images to an average template brain, allowing comparison of brain imaging results. We introduce a novel tool called WarpDrive, which enables manual refinements of image alignment after automated registration. The tool improves accuracy of data representation and aids in understanding patient outcomes.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Ricards Marcinkevics, Patricia Reis Wolfertstetter, Ugne Klimiene, Kieran Chin-Cheong, Alyssia Paschke, Julia Zerres, Markus Denzinger, David Niederberger, Sven Wellmann, Ece Ozkan, Christian Knorr, Julia E. Vogt
Summary: This study presents interpretable machine learning models for predicting the diagnosis, management, and severity of suspected appendicitis using ultrasound images. The proposed models utilize concept bottleneck models (CBM) that facilitate interpretation and intervention by clinicians, without compromising performance or requiring time-consuming image annotation.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Jian-Qing Zheng, Ziyang Wang, Baoru Huang, Ngee Han Lim, Bartlomiej W. Papiez
Summary: This article introduces a new method for medical image registration, which utilizes a separable motion backbone and a residual aligner module to better handle the discontinuous motion of multiple neighboring objects. The proposed method achieves excellent registration results on abdominal CT scans and lung CT scans.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Xiangqiong Wu, Guanghua Tan, Hongxia Luo, Zhilun Chen, Bin Pu, Shengli Li, Kenli Li
Summary: This study develops a user-friendly framework for the automated diagnosis of thyroid nodules in ultrasound videos, simulating the diagnostic workflow of radiologists. By interpreting image characteristics and modeling temporal contextual information, the efficiency and generalizability of the diagnosis can be improved.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Riddhish Bhalodia, Shireen Elhabian, Jadie Adams, Wenzheng Tao, Ladislav Kavan, Ross Whitaker
Summary: This paper introduces DeepSSM, a deep learning-based framework for image-to-shape modeling. By learning the functional mapping from images to low-dimensional shape descriptors, DeepSSM can directly infer statistical representation of anatomy from 3D images. Compared to traditional methods, DeepSSM eliminates the need for heavy manual preprocessing and segmentation, and significantly improves computational time.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Florentin Liebmann, Marco von Atzigen, Dominik Stutz, Julian Wolf, Lukas Zingg, Daniel Suter, Nicola A. Cavalcanti, Laura Leoty, Hooman Esfandiari, Jess G. Snedeker, Martin R. Oswald, Marc Pollefeys, Mazda Farshad, Philipp Furnstahl
Summary: This study presents a marker-less approach for automatic registration and real-time navigation of lumbar spinal fusion surgery using a deep neural network, avoiding radiation exposure and surgical errors. The method was validated on an ex-vivo surgery and a public dataset.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Piyush Tiwary, Kinjawl Bhattacharyya, A. P. Prathosh
Summary: Domain shift refers to the change of distributional characteristics between training and testing datasets, leading to performance drop. For medical image tasks, domain shift can be caused by changes in imaging modalities, devices, and staining mechanisms. Existing approaches based on generative models suffer from training difficulties and lack of diversity. In this paper, the authors propose the use of energy-based models (EBMs) for unpaired image-to-image translation in medical images. The proposed method, called Cycle Consistent Twin EBMs (CCT-EBM), employs a pair of EBMs in the latent space of an Auto-Encoder to ensure translation symmetry and coupling between domains.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Yutong Xie, Jianpeng Zhang, Lingqiao Liu, Hu Wang, Yiwen Ye, Johan Verjans, Yong Xia
Summary: This paper proposes a hybrid pre-training paradigm that combines self-supervised learning and supervised learning to improve the representation quality for medical image segmentation tasks. It introduces a reference task in self-supervised learning and optimizes the model using a gradient matching method. The experimental results demonstrate the effectiveness of this approach on multiple medical image segmentation benchmarks.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Youyi Song, Jing Zou, Kup-Sze Choi, Baiying Lei, Jing Qin
Summary: Cell classification is crucial for intelligent cervical cancer screening, but the variation in cells' appearance and shape poses challenges. A new learning algorithm, worse-case boosting, is proposed to improve classification accuracy for under-represented data. Experimental results demonstrate the effectiveness of this algorithm in two publicly available datasets, achieving a 4% improvement in accuracy.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Sangjoon Park, Eun Sun Lee, Kyung Sook Shin, Jeong Eun Lee, Jong Chul Ye
Summary: The increasing demand for AI systems to monitor human errors and abnormalities in healthcare presents challenges. This study presents a model called Medical X-VL, which is tailored for the medical domain and outperformed current state-of-the-art models in two medical image datasets. The model enables various zero-shot tasks for monitoring AI in the medical domain.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Anna Klimovskaia Susmelj, Berkan Lafci, Firat Ozdemir, Neda Davoudi, Xose Luis Dean-Ben, Fernando Perez-Cruz, Daniel Razansky
Summary: Optoacoustic imaging is a technique that uses optical excitation and ultrasound detection for biological tissue imaging. The quality of the images depends on the extent of tomographic coverage provided by the ultrasound detector arrays. However, full coverage is not always possible due to experimental constraints. The proposed signal domain adaptation network aims to reduce limited-view artifacts in the images.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Srijay Deshpande, Muhammad Dawood, Fayyaz Minhas, Nasir Rajpoot
Summary: In this work, a novel framework called SynCLay is proposed for automated synthesis of histology images based on user-defined cellular layouts. The framework can generate realistic and high-quality histology images with different cellular arrangements, which is helpful for studying the role of cells in the tumor microenvironment. The framework integrates a nuclear segmentation and classification model to refine nuclear structures and generate nuclear masks. Evaluation using quantitative metrics and feedback from pathologists shows that the synthetic images generated by SynCLay have high realism scores and can accurately differentiate between benign and malignant tumors.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Ahmed H. Shahin, An Zhao, Alexander C. Whitehead, Daniel C. Alexander, Joseph Jacob, David Barber
Summary: Survival analysis is a valuable tool in healthcare for predicting the time to specific events. This paper introduces CenTime, a novel approach that directly estimates the time to event. The method performs well with censored data and can be easily integrated with deep learning models. Compared to standard methods, CenTime offers superior performance in predicting event time while maintaining comparable ranking performance.
MEDICAL IMAGE ANALYSIS
(2024)
Article
Computer Science, Artificial Intelligence
Bingyuan Liu, Jose Dolz, Adrian Galdran, Riadh Kobbi, Ismail Ben Ayed
Summary: Most segmentation losses, such as CE and Dice, are variants of the Cross-Entropy or Dice losses. This work provides a theoretical analysis that shows a deeper connection between CE and Dice than previously thought. From a constrained-optimization perspective, both CE and Dice decompose into similar ground-truth matching terms and region-size penalty terms. The analysis uncovers hidden region-size biases: Dice has an intrinsic bias towards extremely imbalanced solutions, while CE implicitly encourages the ground-truth region proportions. Based on this analysis, a principled and simple solution is proposed to explicitly control the region-size bias.
MEDICAL IMAGE ANALYSIS
(2024)