Disentangle, Align and Fuse for Multimodal and Semi-Supervised Image Segmentation

MACTFusion: Lightweight Cross Transformer for Adaptive Multimodal Medical Image Fusion

Multimodal medical image fusion aims to integrate complementary information from different modalities of medical images. Deep learning methods, especially recent vision Transformers, have effectively improved image fusion performance. However, there are limitations for Transformers in image fusion, such as lacks of local feature extraction and cross-modal feature interaction, resulting in insufficient multimodal feature extraction and integration. In addition, the computational cost of Transformers is higher. To address these challenges, in this work, we develop an adaptive cross-modal fusion strategy for unsupervised multimodal medical image fusion. Specifically, we propose a novel lightweight cross Transformer based on cross multi-axis attention mechanism. It includes cross-window attention and cross-grid attention to mine and integrate both local and global interactions of multimodal features. The cross Transformer is further guided by a spatial adaptation fusion module, which allows the model to focus on the most relevant information. Moreover, we design a special feature extraction module that combines multiple gradient residual dense convolutional and Transformer layers to obtain local features from coarse to fine and capture global features. The proposed strategy significantly boosts the fusion performance while minimizing computational costs. Extensive experiments, including clinical brain tumor image fusion, have shown that our model can achieve clearer texture details and better visual quality than other state-of-the-art fusion methods.

Causal recurrent intervention for cross-modal cardiac image segmentation

Cross-modal cardiac image segmentation is essential for cardiac disease analysis. In diagnosis, it enables clinicians to obtain more precise information about cardiac structure or function for potential signs by leveraging specific imaging modalities. For instance, cardiovascular pathologies such as myocardial infarction and congenital heart defects require precise cross-modal characterization to guide clinical decisions. The growing adoption of cross-modal segmentation in clinical research underscores its technical value, yet annotating cardiac images with multiple slices is time-consuming and labor-intensive, making it difficult to meet clinical and deep learning demands. To reduce the need for labels, cross-modal approaches could leverage general knowledge from multiple modalities. However, implementing a cross-modal method remains challenging due to cross-domain confounding. This challenge arises from the intricate effects of modality and view alterations between images, including inconsistent high-dimensional features. The confounding complicates the causality between the observation (image) and the prediction (label), thereby weakening the domain-invariant representation. Existing disentanglement methods face difficulties in addressing the confounding due to the insufficient depiction of the relationship between latent factors. This paper proposes the causal recurrent intervention (CRI) method to overcome the above challenge. It establishes a structural causal model that allows individual domains to maintain causal consistency through interventions. The CRI method integrates diverse high-dimensional variations into a singular causal relationship by embedding image slices into a sequence. This approach further distinguishes stable and dynamic factors from the sequence, subsequently separating the stable factor into modal and view factors and establishing causal connections between them. It then learns the dynamic factor and the view factor from the observation to obtain the label. Experimental results on cross-modal cardiac images of 1697 examples show that the CRI method delivers promising and productive cross-modal cardiac image segmentation performance.

Deep learning for cardiac imaging: focus on myocardial diseases, a narrative review

The integration of computational technologies into cardiology has significantly advanced the diagnosis and management of cardiovascular diseases. Computational cardiology, particularly, through cardiovascular imaging and informatics, enables a precise diagnosis of myocardial diseases utilizing techniques such as echocardiography, cardiac magnetic resonance imaging, and computed tomography. Early-stage disease classification, especially in asymptomatic patients, benefits from these advancements, potentially altering disease progression and improving patient outcomes. Automatic segmentation of myocardial tissue using deep learning (DL) algorithms improves efficiency and consistency in analyzing large patient populations. Radiomic analysis can reveal subtle disease characteristics from medical images and can enhance disease detection, enable patient stratification, and facilitate monitoring of disease progression and treatment response. Radiomic biomarkers have already demonstrated high diagnostic accuracy in distinguishing myocardial pathologies and promise treatment individualization in cardiology, earlier disease detection, and disease monitoring. In this context, this narrative review explores the current state of the art in DL applications in medical imaging (CT, CMR, echocardiography, and SPECT), focusing on automatic segmentation, radiomic feature phenotyping, and prediction of myocardial diseases, while also discussing challenges in integration of DL models in clinical practice.

Multimodal deep learning approaches for precision oncology: a comprehensive review

The burgeoning accumulation of large-scale biomedical data in oncology, alongside significant strides in deep learning (DL) technologies, has established multimodal DL (MDL) as a cornerstone of precision oncology. This review provides an overview of MDL applications in this field, based on an extensive literature survey. In total, 651 articles published before September 2024 are included. We first outline publicly available multimodal datasets that support cancer research. Then, we discuss key DL training methods, data representation techniques, and fusion strategies for integrating multimodal data. The review also examines MDL applications in tumor segmentation, detection, diagnosis, prognosis, treatment selection, and therapy response monitoring. Finally, we critically assess the limitations of current approaches and propose directions for future research. By synthesizing current progress and identifying challenges, this review aims to guide future efforts in leveraging MDL to advance precision oncology.

Style Consistency Unsupervised Domain Adaptation Medical Image Segmentation

Unsupervised domain adaptation medical image segmentation is aimed to segment unlabeled target domain images with labeled source domain images. However, different medical imaging modalities lead to large domain shift between their images, in which well-trained models from one imaging modality often fail to segment images from anothor imaging modality. In this paper, to mitigate domain shift between source domain and target domain, a style consistency unsupervised domain adaptation image segmentation method is proposed. First, a local phase-enhanced style fusion method is designed to mitigate domain shift and produce locally enhanced organs of interest. Second, a phase consistency discriminator is constructed to distinguish the phase consistency of domain-invariant features between source domain and target domain, so as to enhance the disentanglement of the domain-invariant and style encoders and removal of domain-specific features from the domain-invariant encoder. Third, a style consistency estimation method is proposed to obtain inconsistency maps from intermediate synthesized target domain images with different styles to measure the difficult regions, mitigate domain shift between synthesized target domain images and real target domain images, and improve the integrity of interested organs. Fourth, style consistency entropy is defined for target domain images to further improve the integrity of the interested organ by the concentration on the inconsistent regions. Comprehensive experiments have been performed with an in-house dataset and a publicly available dataset. The experimental results have demonstrated the superiority of our framework over state-of-the-art methods.

A review of self-supervised, generative, and few-shot deep learning methods for data-limited magnetic resonance imaging segmentation

Magnetic resonance imaging (MRI) is a ubiquitous medical imaging technology with applications in disease diagnostics, intervention, and treatment planning. Accurate MRI segmentation is critical for diagnosing abnormalities, monitoring diseases, and deciding on a course of treatment. With the advent of advanced deep learning frameworks, fully automated and accurate MRI segmentation is advancing. Traditional supervised deep learning techniques have advanced tremendously, reaching clinical-level accuracy in the field of segmentation. However, these algorithms still require a large amount of annotated data, which is oftentimes unavailable or impractical. One way to circumvent this issue is to utilize algorithms that exploit a limited amount of labeled data. This paper aims to review such state-of-the-art algorithms that use a limited number of annotated samples. We explain the fundamental principles of self-supervised learning, generative models, few-shot learning, and semi-supervised learning and summarize their applications in cardiac, abdomen, and brain MRI segmentation. Throughout this review, we highlight algorithms that can be employed based on the quantity of annotated data available. We also present a comprehensive list of notable publicly available MRI segmentation datasets. To conclude, we discuss possible future directions of the field-including emerging algorithms, such as contrastive language-image pretraining, and potential combinations across the methods discussed-that can further increase the efficacy of image segmentation with limited labels.

Semi-supervised multi-modal medical image segmentation with unified translation

The two major challenges to deep-learning-based medical image segmentation are multi-modality and a lack of expert annotations. Existing semi-supervised segmentation models can mitigate the problem of insufficient annotations by utilizing a small amount of labeled data. However, most of these models are limited to single-modal data and cannot exploit the complementary information from multi-modal medical images. A few semi-supervised multi-modal models have been proposed recently, but they have rigid structures and require additional training steps for each modality. In this work, we propose a novel flexible method, semi-supervised multi-modal medical image segmentation with unified translation (SMSUT), and a unique semi-supervised procedure that can leverage multi-modal information to improve the semi-supervised segmentation performance. Our architecture capitalizes on unified translation to extract complementary information from multi-modal data which compels the network to focus on the disparities and salient features among each modality. Furthermore, we impose constraints on the model at both pixel and feature levels, to cope with the lack of annotation information and the diverse representations within semi-supervised multi-modal data. We introduce a novel training procedure tailored for semi-supervised multi-modal medical image analysis, by integrating the concept of conditional translation. Our method has a remarkable ability for seamless adaptation to varying numbers of distinct modalities in the training data. Experiments show that our model exceeds the semi-supervised segmentation counterparts in the public datasets which proves our network's high-performance capabilities and the transferability of our proposed method. The code of our method will be openly available at https://github.com/Sue1347/SMSUT-MedicalImgSegmentation.

Is Attention all You Need in Medical Image Analysis? A Review

Medical imaging is a key component in clinical diagnosis, treatment planning and clinical trial design, accounting for almost 90% of all healthcare data. CNNs achieved performance gains in medical image analysis (MIA) over the last years. CNNs can efficiently model local pixel interactions and be trained on small-scale MI data. Despite their important advances, typical CNN have relatively limited capabilities in modelling "global" pixel interactions, which restricts their generalisation ability to understand out-of-distribution data with different "global" information. The recent progress of Artificial Intelligence gave rise to Transformers, which can learn global relationships from data. However, full Transformer models need to be trained on large-scale data and involve tremendous computational complexity. Attention and Transformer compartments ("Transf/Attention") which can well maintain properties for modelling global relationships, have been proposed as lighter alternatives of full Transformers. Recently, there is an increasing trend to co-pollinate complementary local-global properties from CNN and Transf/Attention architectures, which led to a new era of hybrid models. The past years have witnessed substantial growth in hybrid CNN-Transf/Attention models across diverse MIA problems. In this systematic review, we survey existing hybrid CNN-Transf/Attention models, review and unravel key architectural designs, analyse breakthroughs, and evaluate current and future opportunities as well as challenges. We also introduced an analysis framework on generalisation opportunities of scientific and clinical impact, based on which new data-driven domain generalisation and adaptation methods can be stimulated.

Semisupervised 3D segmentation of pancreatic tumors in positron emission tomography/computed tomography images using a mutual information minimization and cross-fusion strategy

Background

Accurate segmentation of pancreatic cancer tumors using positron emission tomography/computed tomography (PET/CT) multimodal images is crucial for clinical diagnosis and prognosis evaluation. However, deep learning methods for automated medical image segmentation require a substantial amount of manually labeled data, making it time-consuming and labor-intensive. Moreover, addition or simple stitching of multimodal images leads to redundant information, failing to fully exploit the complementary information of multimodal images. Therefore, we developed a semisupervised multimodal network that leverages limited labeled samples and introduces a cross-fusion and mutual information minimization (MIM) strategy for PET/CT 3D segmentation of pancreatic tumors.

Methods

Our approach combined a cross multimodal fusion (CMF) module with a cross-attention mechanism. The complementary multimodal features were fused to form a multifeature set to enhance the effectiveness of feature extraction while preserving specific features of each modal image. In addition, we designed an MIM module to mitigate redundant high-level modal information and compute the latent loss of PET and CT. Finally, our method employed the uncertainty-aware mean teacher semi-supervised framework to segment regions of interest from PET/CT images using a small amount of labeled data and a large amount of unlabeled data.

Results

We evaluated our combined MIM and CMF semisupervised segmentation network (MIM-CMFNet) on a private dataset of pancreatic cancer, yielding an average Dice coefficient of 73.14%, an average Jaccard index score of 60.56%, and an average 95% Hausdorff distance (95HD) of 6.30 mm. In addition, to verify the broad applicability of our method, we used a public dataset of head and neck cancer, yielding an average Dice coefficient of 68.71%, an average Jaccard index score of 57.72%, and an average 95HD of 7.88 mm.

Conclusions

The experimental results demonstrate the superiority of our MIM-CMFNet over existing semisupervised techniques. Our approach can achieve a performance similar to that of fully supervised segmentation methods while significantly reducing the data annotation cost by 80%, suggesting it is highly practicable for clinical application.

Federated Semi-Supervised Medical Image Segmentation via Prototype-Based Pseudo-Labeling and Contrastive Learning

Existing federated learning works mainly focus on the fully supervised training setting. In realistic scenarios, however, most clinical sites can only provide data without annotations due to the lack of resources or expertise. In this work, we are concerned with the practical yet challenging federated semi-supervised segmentation (FSSS), where labeled data are only with several clients and other clients can just provide unlabeled data. We take an early attempt to tackle this problem and propose a novel FSSS method with prototype-based pseudo-labeling and contrastive learning. First, we transmit a labeled-aggregated model, which is obtained based on prototype similarity, to each unlabeled client, to work together with the global model for debiased pseudo labels generation via a consistency- and entropy-aware selection strategy. Second, we transfer image-level prototypes from labeled datasets to unlabeled clients and conduct prototypical contrastive learning on unlabeled models to enhance their discriminative power. Finally, we perform the dynamic model aggregation with a designed consistency-aware aggregation strategy to dynamically adjust the aggregation weights of each local model. We evaluate our method on COVID-19 X-ray infected region segmentation, COVID-19 CT infected region segmentation and colorectal polyp segmentation, and experimental results consistently demonstrate the effectiveness of our proposed method. Codes areavailable at https://github.com/zhangbaiming/FedSemiSeg.

Applications of AI in multi-modal imaging for cardiovascular disease

Data for healthcare is diverse and includes many different modalities. Traditional approaches to Artificial Intelligence for cardiovascular disease were typically limited to single modalities. With the proliferation of diverse datasets and new methods in AI, we are now able to integrate different modalities, such as magnetic resonance scans, computerized tomography scans, echocardiography, x-rays, and electronic health records. In this paper, we review research from the last 5 years in applications of AI to multi-modal imaging. There have been many promising results in registration, segmentation, and fusion of different magnetic resonance imaging modalities with each other and computer tomography scans, but there are still many challenges that need to be addressed. Only a few papers have addressed modalities such as x-ray, echocardiography, or non-imaging modalities. As for prediction or classification tasks, there have only been a couple of papers that use multiple modalities in the cardiovascular domain. Furthermore, no models have been implemented or tested in real world cardiovascular clinical settings.

Unsupervised Domain Adaptation for Medical Image Segmentation by Disentanglement Learning and Self-Training

Unsupervised domain adaption (UDA), which aims to enhance the segmentation performance of deep models on unlabeled data, has recently drawn much attention. In this paper, we propose a novel UDA method (namely DLaST) for medical image segmentation via disentanglement learning and self-training. Disentanglement learning factorizes an image into domain-invariant anatomy and domain-specific modality components. To make the best of disentanglement learning, we propose a novel shape constraint to boost the adaptation performance. The self-training strategy further adaptively improves the segmentation performance of the model for the target domain through adversarial learning and pseudo label, which implicitly facilitates feature alignment in the anatomy space. Experimental results demonstrate that the proposed method outperforms the state-of-the-art UDA methods for medical image segmentation on three public datasets, i.e., a cardiac dataset, an abdominal dataset and a brain dataset. The code will be released soon.

Learning with limited target data to detect cells in cross-modality images

Deep neural networks have achieved excellent cell or nucleus quantification performance in microscopy images, but they often suffer from performance degradation when applied to cross-modality imaging data. Unsupervised domain adaptation (UDA) based on generative adversarial networks (GANs) has recently improved the performance of cross-modality medical image quantification. However, current GAN-based UDA methods typically require abundant target data for model training, which is often very expensive or even impossible to obtain for real applications. In this paper, we study a more realistic yet challenging UDA situation, where (unlabeled) target training data is limited and previous work seldom delves into cell identification. We first enhance a dual GAN with task-specific modeling, which provides additional supervision signals to assist with generator learning. We explore both single-directional and bidirectional task-augmented GANs for domain adaptation. Then, we further improve the GAN by introducing a differentiable, stochastic data augmentation module to explicitly reduce discriminator overfitting. We examine source-, target-, and dual-domain data augmentation for GAN enhancement, as well as joint task and data augmentation in a unified GAN-based UDA framework. We evaluate the framework for cell detection on multiple public and in-house microscopy image datasets, which are acquired with different imaging modalities, staining protocols and/or tissue preparations. The experiments demonstrate that our method significantly boosts performance when compared with the reference baseline, and it is superior to or on par with fully supervised models that are trained with real target annotations. In addition, our method outperforms recent state-of-the-art UDA approaches by a large margin on different datasets.

Shape-aware Segmentation of the Placenta in BOLD Fetal MRI Time Series

Blood oxygen level dependent (BOLD) MRI time series with maternal hyperoxia can assess placental oxygenation and function. Measuring precise BOLD changes in the placenta requires accurate temporal placental segmentation and is confounded by fetal and maternal motion, contractions, and hyperoxia-induced intensity changes. Current BOLD placenta segmentation methods warp a manually annotated subject-specific template to the entire time series. However, as the placenta is a thin, elongated, and highly non-rigid organ subject to large deformations and obfuscated edges, existing work cannot accurately segment the placental shape, especially near boundaries. In this work, we propose a machine learning segmentation framework for placental BOLD MRI and apply it to segmenting each volume in a time series. We use a placental-boundary weighted loss formulation and perform a comprehensive evaluation across several popular segmentation objectives. Our model is trained and tested on a cohort of 91 subjects containing healthy fetuses, fetuses with fetal growth restriction, and mothers with high BMI. Biomedically, our model performs reliably in segmenting volumes in both normoxic and hyperoxic points in the BOLD time series. We further find that boundary-weighting increases placental segmentation performance by 8.3% and 6.0% Dice coefficient for the cross-entropy and signed distance transform objectives, respectively.

CDDSA: Contrastive domain disentanglement and style augmentation for generalizable medical image segmentation

Generalization to previously unseen images with potential domain shifts is essential for clinically applicable medical image segmentation. Disentangling domain-specific and domain-invariant features is key for Domain Generalization (DG). However, existing DG methods struggle to achieve effective disentanglement. To address this problem, we propose an efficient framework called Contrastive Domain Disentanglement and Style Augmentation (CDDSA) for generalizable medical image segmentation. First, a disentangle network decomposes the image into domain-invariant anatomical representation and domain-specific style code, where the former is sent for further segmentation that is not affected by domain shift, and the disentanglement is regularized by a decoder that combines the anatomical representation and style code to reconstruct the original image. Second, to achieve better disentanglement, a contrastive loss is proposed to encourage the style codes from the same domain and different domains to be compact and divergent, respectively. Finally, to further improve generalizability, we propose a style augmentation strategy to synthesize images with various unseen styles in real time while maintaining anatomical information. Comprehensive experiments on a public multi-site fundus image dataset and an in-house multi-site Nasopharyngeal Carcinoma Magnetic Resonance Image (NPC-MRI) dataset show that the proposed CDDSA achieved remarkable generalizability across different domains, and it outperformed several state-of-the-art methods in generalizable segmentation. Code is available at https://github.com/HiLab-git/DAG4MIA.

DE-JANet: A unified network based on dual encoder and joint attention for Alzheimer's disease classification using multi-modal data

Structural magnetic resonance imaging (sMRI), which can reflect cerebral atrophy, plays an important role in the early detection of Alzheimer's disease (AD). However, the information provided by analyzing only the morphological changes in sMRI is relatively limited, and the assessment of the atrophy degree is subjective. Therefore, it is meaningful to combine sMRI with other clinical information to acquire complementary diagnosis information and achieve a more accurate classification of AD. Nevertheless, how to fuse these multi-modal data effectively is still challenging. In this paper, we propose DE-JANet, a unified AD classification network that integrates image data sMRI with non-image clinical data, such as age and Mini-Mental State Examination (MMSE) score, for more effective multi-modal analysis. DE-JANet consists of three key components: (1) a dual encoder module for extracting low-level features from the image and non-image data according to specific encoding regularity, (2) a joint attention module for fusing multi-modal features, and (3) a token classification module for performing AD-related classification according to the fused multi-modal features. Our DE-JANet is evaluated on the ADNI dataset, with a mean accuracy of 0.9722 and 0.9538 for AD classification and mild cognition impairment (MCI) classification, respectively, which is superior to existing methods and indicates advanced performance on AD-related diagnosis tasks.

Multi-modality cardiac image computing: A survey

Multi-modality cardiac imaging plays a key role in the management of patients with cardiovascular diseases. It allows a combination of complementary anatomical, morphological and functional information, increases diagnosis accuracy, and improves the efficacy of cardiovascular interventions and clinical outcomes. Fully-automated processing and quantitative analysis of multi-modality cardiac images could have a direct impact on clinical research and evidence-based patient management. However, these require overcoming significant challenges including inter-modality misalignment and finding optimal methods to integrate information from different modalities. This paper aims to provide a comprehensive review of multi-modality imaging in cardiology, the computing methods, the validation strategies, the related clinical workflows and future perspectives. For the computing methodologies, we have a favored focus on the three tasks, i.e., registration, fusion and segmentation, which generally involve multi-modality imaging data, either combining information from different modalities or transferring information across modalities. The review highlights that multi-modality cardiac imaging data has the potential of wide applicability in the clinic, such as trans-aortic valve implantation guidance, myocardial viability assessment, and catheter ablation therapy and its patient selection. Nevertheless, many challenges remain unsolved, such as missing modality, modality selection, combination of imaging and non-imaging data, and uniform analysis and representation of different modalities. There is also work to do in defining how the well-developed techniques fit in clinical workflows and how much additional and relevant information they introduce. These problems are likely to continue to be an active field of research and the questions to be answered in the future.

Lazy Luna: Extendible software for multilevel reader comparison in cardiovascular magnetic resonance imaging

BACKGROUND AND OBJECTIVES

Cardiovascular Magnetic Resonance (CMR) imaging is a growing field with increasing diagnostic utility in clinical routine. Quantitative diagnostic parameters are typically calculated based on contours or points provided by readers, e.g. natural intelligences (NI) such as clinicians or researchers, and artificial intelligences (AI). As clinical applications multiply, evaluating the precision and reproducibility of quantitative parameters becomes increasingly important. Although segmentation challenges for AIs and guidelines for clinicians provide quality assessments and regulation, the methods ought to be combined and streamlined for clinical applications. The goal of the developed software, Lazy Luna (LL), is to offer a flexible evaluation tool that is readily extendible to new sequences and scientific endeavours.

METHODS

An interface was designed for LL, which allows for comparing annotated CMR images. Geometric objects ensure precise calculations of metric values and clinical results regardless of whether annotations originate from AIs or NIs. A graphical user interface (GUI) is provided to make the software available to non-programmers. The GUI allows for an interactive inspection of image datasets as well as implementing tracing procedures, which follow statistical reader differences in clinical results to their origins in individual image contours. The backend software builds on a set of meta-classes, which can be extended to new imaging sequences and clinical parameters. Following an agile development procedure with clinical feedback allows for a quick implementation of new classes, figures and tables for evaluation.

RESULTS

Two application cases present LL's extendibility to clinical evaluation and AI development contexts. The first concerns T1 parametric mapping images segmented by two expert readers. Quantitative result differences are traced to reveal typical segmentation dissimilarities from which these differences originate. The meta-classes are extended to this new application scenario. The second applies to the open source Late Gadolinium Enhancement (LGE) quantification challenge for AI developers "Emidec", which illustrates LL's usability as open source software.

CONCLUSION

The presented software Lazy Luna allows for an automated multilevel comparison of readers as well as identifying qualitative reasons for statistical reader differences. The open source software LL can be extended to new application cases in the future.

MDST: multi-domain sparse-view CT reconstruction based on convolution and swin transformer

Objective.Sparse-view computed tomography (SVCT), which can reduce the radiation doses administered to patients and hasten data acquisition, has become an area of particular interest to researchers. Most existing deep learning-based image reconstruction methods are based on convolutional neural networks (CNNs). Due to the locality of convolution and continuous sampling operations, existing approaches cannot fully model global context feature dependencies, which makes the CNN-based approaches less efficient in modeling the computed tomography (CT) images with various structural information.Approach.To overcome the above challenges, this paper develops a novel multi-domain optimization network based on convolution and swin transformer (MDST). MDST uses swin transformer block as the main building block in both projection (residual) domain and image (residual) domain sub-networks, which models global and local features of the projections and reconstructed images. MDST consists of two modules for initial reconstruction and residual-assisted reconstruction, respectively. The sparse sinogram is first expanded in the initial reconstruction module with a projection domain sub-network. Then, the sparse-view artifacts are effectively suppressed by an image domain sub-network. Finally, the residual assisted reconstruction module to correct the inconsistency of the initial reconstruction, further preserving image details.Main results. Extensive experiments on CT lymph node datasets and real walnut datasets show that MDST can effectively alleviate the loss of fine details caused by information attenuation and improve the reconstruction quality of medical images.Significance.MDST network is robust and can effectively reconstruct images with different noise level projections. Different from the current prevalent CNN-based networks, MDST uses transformer as the main backbone, which proves the potential of transformer in SVCT reconstruction.

Learning multi-modal brain tumor segmentation from privileged semi-paired MRI images with curriculum disentanglement learning

Since the brain is the human body's primary command and control center, brain cancer is one of the most dangerous cancers. Automatic segmentation of brain tumors from multi-modal images is important in diagnosis and treatment. Due to the difficulties in obtaining multi-modal paired images in clinical practice, recent studies segment brain tumors solely relying on unpaired images and discarding the available paired images. Although these models solve the dependence on paired images, they cannot fully exploit the complementary information from different modalities, resulting in low unimodal segmentation accuracy. Hence, this work studies the unimodal segmentation with privileged semi-paired images, i.e., limited paired images are introduced to the training phase. Specifically, we present a novel two-step (intra-modality and inter-modality) curriculum disentanglement learning framework. The modality-specific style codes describe the attenuation of tissue features and image contrast, and modality-invariant content codes contain anatomical and functional information extracted from the input images. Besides, we address the problem of unthorough decoupling by introducing constraints on the style and content spaces. Experiments on the BraTS2020 dataset highlight that our model outperforms the competing models on unimodal segmentation, achieving average dice scores of 82.91%, 72.62%, and 54.80% for WT (the whole tumor), TC (the tumor core), and ET (the enhancing tumor), respectively. Finally, we further evaluate our model's variable multi-modal brain tumor segmentation performance by introducing a fusion block (TFusion). The experimental results reveal that our model achieves the best WT segmentation performance for all 15 possible modality combinations with 87.31% average accuracy. In summary, we propose a curriculum disentanglement learning framework for unimodal segmentation with privileged semi-paired images. Moreover, the benefits of the improved unimodal segmentation extend to variable multi-modal segmentation, demonstrating that improving the unimodal segmentation performance is significant for brain tumor segmentation with missing modalities. Our code is available at https://github.com/scut-cszcl/SpBTS.

Multilevel comparison of deep learning models for function quantification in cardiovascular magnetic resonance: On the redundancy of architectural variations

Background

Cardiac function quantification in cardiovascular magnetic resonance requires precise contouring of the heart chambers. This time-consuming task is increasingly being addressed by a plethora of ever more complex deep learning methods. However, only a small fraction of these have made their way from academia into clinical practice. In the quality assessment and control of medical artificial intelligence, the opaque reasoning and associated distinctive errors of neural networks meet an extraordinarily low tolerance for failure.

Aim

The aim of this study is a multilevel analysis and comparison of the performance of three popular convolutional neural network (CNN) models for cardiac function quantification.

Methods

U-Net, FCN, and MultiResUNet were trained for the segmentation of the left and right ventricles on short-axis cine images of 119 patients from clinical routine. The training pipeline and hyperparameters were kept constant to isolate the influence of network architecture. CNN performance was evaluated against expert segmentations for 29 test cases on contour level and in terms of quantitative clinical parameters. Multilevel analysis included breakdown of results by slice position, as well as visualization of segmentation deviations and linkage of volume differences to segmentation metrics via correlation plots for qualitative analysis.

Results

All models showed strong correlation to the expert with respect to quantitative clinical parameters (rz ' = 0.978, 0.977, 0.978 for U-Net, FCN, MultiResUNet respectively). The MultiResUNet significantly underestimated ventricular volumes and left ventricular myocardial mass. Segmentation difficulties and failures clustered in basal and apical slices for all CNNs, with the largest volume differences in the basal slices (mean absolute error per slice: 4.2 ± 4.5 ml for basal, 0.9 ± 1.3 ml for midventricular, 0.9 ± 0.9 ml for apical slices). Results for the right ventricle had higher variance and more outliers compared to the left ventricle. Intraclass correlation for clinical parameters was excellent (≥0.91) among the CNNs.

Conclusion

Modifications to CNN architecture were not critical to the quality of error for our dataset. Despite good overall agreement with the expert, errors accumulated in basal and apical slices for all models.

NVTrans-UNet: Neighborhood vision transformer based U-Net for multi-modal cardiac MR image segmentation

With the rapid development of artificial intelligence and image processing technology, medical imaging technology has turned into a critical tool for clinical diagnosis and disease treatment. The extraction and segmentation of the regions of interest in cardiac images are crucial to the diagnosis of cardiovascular diseases. Due to the erratically diastolic and systolic cardiac, the boundaries of Magnetic Resonance (MR) images are quite fuzzy. Moreover, it is hard to provide complete information using a single modality due to the complex structure of the cardiac image. Furthermore, conventional CNN-based segmentation methods are weak in feature extraction. To overcome these challenges, we propose a multi-modal method for cardiac image segmentation, called NVTrans-UNet. Firstly, we employ the Neighborhood Vision Transformer (NVT) module, which takes advantage of Neighborhood Attention (NA) and inductive biases. It can better extract the local information of the cardiac image as well as reduce the computational cost. Secondly, we introduce a Multi-modal Gated Fusion (MGF) network, which can automatically adjust the contributions of different modal feature maps and make full use of multi-modal information. Thirdly, the bottleneck layer with Atrous Spatial Pyramid Pooling (ASPP) is proposed to expand the feature receptive field. Finally, the mixed loss is added to the cardiac image to focus the fuzzy boundary and realize accurate segmentation. We evaluated our model on MyoPS 2020 dataset. The Dice score of myocardial infarction (MI) was 0.642 ± 0.171, and the Dice score of myocardial infarction + edema (MI + ME) was 0.574 ± 0.110. Compared with the baseline, the MI increases by 11.2%, and the MI + ME increases by 12.5%. The results show the effectiveness of the proposed NVTrans-UNet in the segmentation of MI and ME.

HD2A-Net: A novel dual gated attention network using comprehensive hybrid dilated convolutions for medical image segmentation

The convolutional neural networks (CNNs) have been widely proposed in the medical image analysis tasks, especially in the image segmentations. In recent years, the encoder-decoder structures, such as the U-Net, were rendered. However, the multi-scale information transmission and effective modeling for long-range feature dependencies in these structures were not sufficiently considered. To improve the performance of the existing methods, we propose a novel hybrid dual dilated attention network (HD²A-Net) to conduct the lesion region segmentations. In the proposed network, we innovatively present the comprehensive hybrid dilated convolution (CHDC) module, which facilitates the transmission of the multi-scale information. Based on the CHDC module and the attention mechanisms, we design a novel dual dilated gated attention (DDGA) block to enhance the saliency of related regions from the multi-scale aspect. Besides, a dilated dense (DD) block is designed to expand the receptive fields. The ablation studies were performed to verify our proposed blocks. Besides, the interpretability of the HD²A-Net was analyzed through the visualization of the attention weight maps from the key blocks. Compared to the state-of-the-art methods including CA-Net, DeepLabV3+, and Attention U-Net, the HD²A-Net outperforms significantly, with the metrics of Dice, Average Symmetric Surface Distance (ASSD), and mean Intersection-over-Union (mIoU) reaching 93.16%, 93.63%, and 94.72%, 0.36 pix, 0.69 pix, and 0.52 pix, and 88.03%, 88.67%, and 90.33% on three publicly available medical image datasets: MAEDE-MAFTOUNI (COVID-19 CT), ISIC-2018 (Melanoma Dermoscopy), and Kvasir-SEG (Gastrointestinal Disease Polyp), respectively.

Motion correction for native myocardial T1 mapping using self-supervised deep learning registration with contrast separation

In myocardial T₁ mapping, undesirable motion poses significant challenges because uncorrected motion can affect T₁ estimation accuracy and cause incorrect diagnosis. In this study, we propose and evaluate a motion correction method for myocardial T₁ mapping using self-supervised deep learning based registration with contrast separation (SDRAP). A sparse coding based method was first proposed to separate the contrast component from T₁ -weighted (T1w) images. Then, a self-supervised deep neural network with cross-correlation (SDRAP-CC) or mutual information as the registration similarity measurement was developed to register contrast separated images, after which signal fitting was performed on the motion corrected T1w images to generate motion corrected T₁ maps. The registration network was trained and tested in 80 healthy volunteers with images acquired using the modified Look-Locker inversion recovery (MOLLI) sequence. The proposed SDRAP was compared with the free form deformation (FFD) registration method regarding (1) Dice similarity coefficient (DSC) and mean boundary error (MBE) of myocardium contours, (2) T₁ value and standard deviation (SD) of T₁ fitting, (3) subjective evaluation score for overall image quality and motion artifact level, and (4) computation time. Results showed that SDRAP-CC achieved the highest DSC of 85.0 ± 3.9% and the lowest MBE of 0.92 ± 0.25 mm among the methods compared. Additionally, SDRAP-CC performed the best by resulting in lower SD value (28.1 ± 17.6 ms) and higher subjective image quality scores (3.30 ± 0.79 for overall quality and 3.53 ± 0.68 for motion artifact) evaluated by a cardiologist. The proposed SDRAP took only 0.52 s to register one slice of MOLLI images, achieving about sevenfold acceleration over FFD (3.7 s/slice).

Learning disentangled representations in the imaging domain

Disentangled representation learning has been proposed as an approach to learning general representations even in the absence of, or with limited, supervision. A good general representation can be fine-tuned for new target tasks using modest amounts of data, or used directly in unseen domains achieving remarkable performance in the corresponding task. This alleviation of the data and annotation requirements offers tantalising prospects for applications in computer vision and healthcare. In this tutorial paper, we motivate the need for disentangled representations, revisit key concepts, and describe practical building blocks and criteria for learning such representations. We survey applications in medical imaging emphasising choices made in exemplar key works, and then discuss links to computer vision applications. We conclude by presenting limitations, challenges, and opportunities.

A Local and Global Feature Disentangled Network: Toward Classification of Benign-Malignant Thyroid Nodules From Ultrasound Image

Thyroid nodules are one of the most common nodular lesions. The incidence of thyroid cancer has increased rapidly in the past three decades and is one of the cancers with the highest incidence. As a non-invasive imaging modality, ultrasonography can identify benign and malignant thyroid nodules, and it can be used for large-scale screening. In this study, inspired by the domain knowledge of sonographers when diagnosing ultrasound images, a local and global feature disentangled network (LoGo-Net) is proposed to classify benign and malignant thyroid nodules. This model imitates the dual-pathway structure of human vision and establishes a new feature extraction method to improve the recognition performance of nodules. We use the tissue-anatomy disentangled (TAD) block to connect the dual pathways, which decouples the cues of local and global features based on the self-attention mechanism. To verify the effectiveness of the model, we constructed a large-scale dataset and conducted extensive experiments. The results show that our method achieves an accuracy of 89.33%, which has the potential to be used in the clinical practice of doctors, including early cancer screening procedures in remote or resource-poor areas.

Unsupervised Image Registration towards Enhancing Performance and Explainability in Cardiac and Brain Image Analysis

Magnetic Resonance Imaging (MRI) typically recruits multiple sequences (defined here as "modalities"). As each modality is designed to offer different anatomical and functional clinical information, there are evident disparities in the imaging content across modalities. Inter- and intra-modality affine and non-rigid image registration is an essential medical image analysis process in clinical imaging, as for example before imaging biomarkers need to be derived and clinically evaluated across different MRI modalities, time phases and slices. Although commonly needed in real clinical scenarios, affine and non-rigid image registration is not extensively investigated using a single unsupervised model architecture. In our work, we present an unsupervised deep learning registration methodology that can accurately model affine and non-rigid transformations, simultaneously. Moreover, inverse-consistency is a fundamental inter-modality registration property that is not considered in deep learning registration algorithms. To address inverse consistency, our methodology performs bi-directional cross-modality image synthesis to learn modality-invariant latent representations, and involves two factorised transformation networks (one per each encoder-decoder channel) and an inverse-consistency loss to learn topology-preserving anatomical transformations. Overall, our model (named "FIRE") shows improved performances against the reference standard baseline method (i.e., Symmetric Normalization implemented using the ANTs toolbox) on multi-modality brain 2D and 3D MRI and intra-modality cardiac 4D MRI data experiments. We focus on explaining model-data components to enhance model explainability in medical image registration. On computational time experiments, we show that the FIRE model performs on a memory-saving mode, as it can inherently learn topology-preserving image registration directly in the training phase. We therefore demonstrate an efficient and versatile registration technique that can have merit in multi-modal image registrations in the clinical setting.

Decomposing normal and abnormal features of medical images for content-based image retrieval of glioma imaging

In medical imaging, the characteristics purely derived from a disease should reflect the extent to which abnormal findings deviate from the normal features. Indeed, physicians often need corresponding images without abnormal findings of interest or, conversely, images that contain similar abnormal findings regardless of normal anatomical context. This is called comparative diagnostic reading of medical images, which is essential for a correct diagnosis. To support comparative diagnostic reading, content-based image retrieval (CBIR) that can selectively utilize normal and abnormal features in medical images as two separable semantic components will be useful. In this study, we propose a neural network architecture to decompose the semantic components of medical images into two latent codes: normal anatomy code and abnormal anatomy code. The normal anatomy code represents counterfactual normal anatomies that should have existed if the sample is healthy, whereas the abnormal anatomy code attributes to abnormal changes that reflect deviation from the normal baseline. By calculating the similarity based on either normal or abnormal anatomy codes or the combination of the two codes, our algorithm can retrieve images according to the selected semantic component from a dataset consisting of brain magnetic resonance images of gliomas. Moreover, it can utilize a synthetic query vector combining normal and abnormal anatomy codes from two different query images. To evaluate whether the retrieved images are acquired according to the targeted semantic component, the overlap of the ground-truth labels is calculated as metrics of the semantic consistency. Our algorithm provides a flexible CBIR framework by handling the decomposed features with qualitatively and quantitatively remarkable results.

Representation Disentanglement for Multi-modal Brain MRI Analysis

Multi-modal MRIs are widely used in neuroimaging applications since different MR sequences provide complementary information about brain structures. Recent works have suggested that multi-modal deep learning analysis can benefit from explicitly disentangling anatomical (shape) and modality (appearance) information into separate image presentations. In this work, we challenge mainstream strategies by showing that they do not naturally lead to representation disentanglement both in theory and in practice. To address this issue, we propose a margin loss that regularizes the similarity in relationships of the representations across subjects and modalities. To enable robust training, we further use a conditional convolution to design a single model for encoding images of all modalities. Lastly, we propose a fusion function to combine the disentangled anatomical representations as a set of modality-invariant features for downstream tasks. We evaluate the proposed method on three multi-modal neuroimaging datasets. Experiments show that our proposed method can achieve superior disentangled representations compared to existing disentanglement strategies. Results also indicate that the fused anatomical representation has potential in the downstream task of zero-dose PET reconstruction and brain tumor segmentation.