Medical image computing and computer-assisted intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention最新文献

Magnetic resonance imaging (MRI) and positron emission tomography (PET) are increasingly used to forecast progression trajectories of cognitive decline caused by preclinical and prodromal Alzheimer's disease (AD). Many existing studies have explored the potential of these two distinct modalities with diverse machine and deep learning approaches. But successfully fusing MRI and PET can be complex due to their unique characteristics and missing modalities. To this end, we develop a hybrid multimodality fusion (HMF) framework with cross-domain knowledge transfer for joint MRI and PET representation learning, feature fusion, and cognitive decline progression forecasting. Our HMF consists of three modules: 1) a module to impute missing PET images, 2) a module to extract multimodality features from MRI and PET images, and 3) a module to fuse the extracted multimodality features. To address the issue of small sample sizes, we employ a cross-domain knowledge transfer strategy from the ADNI dataset, which includes 795 subjects, to independent small-scale AD-related cohorts, in order to leverage the rich knowledge present within the ADNI. The proposed HMF is extensively evaluated in three AD-related studies with 272 subjects across multiple disease stages, such as subjective cognitive decline and mild cognitive impairment. Experimental results demonstrate the superiority of our method over several state-of-the-art approaches in forecasting progression trajectories of AD-related cognitive decline.

Capturing global contextual information plays a critical role in breast ultrasound (BUS) image classification. Although convolutional neural networks (CNNs) have demonstrated reliable performance in tumor classification, they have inherent limitations for modeling global and long-range dependencies due to the localized nature of convolution operations. Vision Transformers have an improved capability of capturing global contextual information but may distort the local image patterns due to the tokenization operations. In this study, we proposed a hybrid multitask deep neural network called Hybrid-MT-ESTAN, designed to perform BUS tumor classification and segmentation using a hybrid architecture composed of CNNs and Swin Transformer components. The proposed approach was compared to nine BUS classification methods and evaluated using seven quantitative metrics on a dataset of 3,320 BUS images. The results indicate that Hybrid-MT-ESTAN achieved the highest accuracy, sensitivity, and F1 score of 82.7%, 86.4%, and 86.0%, respectively.

We propose an unsupervised deep learning algorithm for the motion-compensated reconstruction of 5D cardiac MRI data from 3D radial acquisitions. Ungated free-breathing 5D MRI simplifies the scan planning, improves patient comfort, and offers several clinical benefits over breath-held 2D exams, including isotropic spatial resolution and the ability to reslice the data to arbitrary views. However, the current reconstruction algorithms for 5D MRI take very long computational time, and their outcome is greatly dependent on the uniformity of the binning of the acquired data into different physiological phases. The proposed algorithm is a more data-efficient alternative to current motion-resolved reconstructions. This motion-compensated approach models the data in each cardiac/respiratory bin as Fourier samples of the deformed version of a 3D image template. The deformation maps are modeled by a convolutional neural network driven by the physiological phase information. The deformation maps and the template are then jointly estimated from the measured data. The cardiac and respiratory phases are estimated from 1D navigators using an auto-encoder. The proposed algorithm is validated on 5D bSSFP datasets acquired from two subjects.

The accuracy of predictive models for solitary pulmonary nodule (SPN) diagnosis can be greatly increased by incorporating repeat imaging and medical context, such as electronic health records (EHRs). However, clinically routine modalities such as imaging and diagnostic codes can be asynchronous and irregularly sampled over different time scales which are obstacles to longitudinal multimodal learning. In this work, we propose a transformer-based multimodal strategy to integrate repeat imaging with longitudinal clinical signatures from routinely collected EHRs for SPN classification. We perform unsupervised disentanglement of latent clinical signatures and leverage time-distance scaled self-attention to jointly learn from clinical signatures expressions and chest computed tomography (CT) scans. Our classifier is pretrained on 2,668 scans from a public dataset and 1,149 subjects with longitudinal chest CTs, billing codes, medications, and laboratory tests from EHRs of our home institution. Evaluation on 227 subjects with challenging SPNs revealed a significant AUC improvement over a longitudinal multimodal baseline (0.824 vs 0.752 AUC), as well as improvements over a single cross-section multimodal scenario (0.809 AUC) and a longitudinal imaging-only scenario (0.741 AUC). This work demonstrates significant advantages with a novel approach for co-learning longitudinal imaging and non-imaging phenotypes with transformers. Code available at https://github.com/MASILab/lmsignatures.

Brain structural MRI has been widely used for assessing future progression of cognitive impairment (CI) based on learning-based methods. Previous studies generally suffer from the limited number of labeled training data, while there exists a huge amount of MRIs in large-scale public databases. Even without task-specific label information, brain anatomical structures provided by these MRIs can be used to boost learning performance intuitively. Unfortunately, existing research seldom takes advantage of such brain anatomy prior. To this end, this paper proposes a brain anatomy-guided representation (BAR) learning framework for assessing the clinical progression of cognitive impairment with T1-weighted MRIs. The BAR consists of a pretext model and a downstream model, with a shared brain anatomy-guided encoder for MRI feature extraction. The pretext model also contains a decoder for brain tissue segmentation, while the downstream model relies on a predictor for classification. We first train the pretext model through a brain tissue segmentation task on 9,544 auxiliary T1-weighted MRIs, yielding a generalizable encoder. The downstream model with the learned encoder is further fine-tuned on target MRIs for prediction tasks. We validate the proposed BAR on two CI-related studies with a total of 391 subjects with T1-weighted MRIs. Experimental results suggest that the BAR outperforms several state-of-the-art (SOTA) methods. The source code and pre-trained models are available at https://github.com/goodaycoder/BAR.

The fiber orientation distribution function (FOD) is an advanced model for high angular resolution diffusion MRI representing complex fiber geometry. However, the complicated mathematical structures of the FOD function pose challenges for FOD image processing tasks such as interpolation, which plays a critical role in the propagation of fiber tracts in tractography. In FOD-based tractography, linear interpolation is commonly used for numerical efficiency, but it is prone to generate false artificial information, leading to anatomically incorrect fiber tracts. To overcome this difficulty, we propose a flowbased and geometrically consistent interpolation framework that considers peak-wise rotations of FODs within the neighborhood of each location. Our method decomposes a FOD function into multiple components and uses a smooth vector field to model the flows of each peak in its neighborhood. To generate the interpolated result along the flow of each vector field, we develop a closed-form and efficient method to rotate FOD peaks in neighboring voxels and realize geometrically consistent interpolation of FOD components. By combining the interpolation results from each peak, we obtain the final interpolation of FODs. Experimental results on Human Connectome Project (HCP) data demonstrate that our method produces anatomically more meaningful FOD interpolations and significantly enhances tractography performance.

Deep learning nowadays offers expert-level and sometimes even super-expert-level performance, but achieving such performance demands massive annotated data for training (e.g., Google's proprietary CXR Foundation Model (CXR-FM) was trained on 821,544 labeled and mostly private chest X-rays (CXRs)). Numerous datasets are publicly available in medical imaging but individually small and heterogeneous in expert labels. We envision a powerful and robust foundation model that can be trained by aggregating numerous small public datasets. To realize this vision, we have developed Ark, a framework that accrues and reuses knowledge from heterogeneous expert annotations in various datasets. As a proof of concept, we have trained two Ark models on 335,484 and 704,363 CXRs, respectively, by merging several datasets including ChestX-ray14, CheXpert, MIMIC-II, and VinDr-CXR, evaluated them on a wide range of imaging tasks covering both classification and segmentation via fine-tuning, linear-probing, and gender-bias analysis, and demonstrated our Ark's superior and robust performance over the state-of-the-art (SOTA) fully/self-supervised baselines and Google's proprietary CXR-FM. This enhanced performance is attributed to our simple yet powerful observation that aggregating numerous public datasets diversifies patient populations and accrues knowledge from diverse experts, yielding unprecedented performance yet saving annotation cost. With all codes and pretrained models released at GitHub.com/JLiangLab/Ark, we hope that Ark exerts an important impact on open science, as accruing and reusing knowledge from expert annotations in public datasets can potentially surpass the performance of proprietary models trained on unusually large data, inspiring many more researchers worldwide to share codes and datasets to build open foundation models, accelerate open science, and democratize deep learning for medical imaging.

One of the hallmark symptoms of Parkinson's Disease (PD) is the progressive loss of postural reflexes, which eventually leads to gait difficulties and balance problems. Identifying disruptions in brain function associated with gait impairment could be crucial in better understanding PD motor progression, thus advancing the development of more effective and personalized therapeutics. In this work, we present an explainable, geometric, weighted-graph attention neural network (xGW-GAT) to identify functional networks predictive of the progression of gait difficulties in individuals with PD. xGW-GAT predicts the multi-class gait impairment on the MDS-Unified PD Rating Scale (MDS-UPDRS). Our computational- and data-efficient model represents functional connectomes as symmetric positive definite (SPD) matrices on a Riemannian manifold to explicitly encode pairwise interactions of entire connectomes, based on which we learn an attention mask yielding individual- and group-level explainability. Applied to our resting-state functional MRI (rs-fMRI) dataset of individuals with PD, xGW-GAT identifies functional connectivity patterns associated with gait impairment in PD and offers interpretable explanations of functional subnetworks associated with motor impairment. Our model successfully outperforms several existing methods while simultaneously revealing clinically-relevant connectivity patterns. The source code is available at https://github.com/favour-nerrise/xGW-GAT.

Deep learning has drawn increasing attention in microstructure estimation with undersampled diffusion MRI (dMRI) data. A representative method is the hybrid graph transformer (HGT), which achieves promising performance by integrating $q$ -space graph learning and $x$ -space transformer learning into a unified framework. However, this method overlooks the 3D spatial information as it relies on training with 2D slices. To address this limitation, we propose 3D hybrid graph transformer (3D-HGT), an advanced microstructure estimation model capable of making full use of 3D spatial information and angular information. To tackle the large computation burden associated with 3D $x$ -space learning, we propose an efficient $q$ -space learning model based on simplified graph neural networks. Furthermore, we propose a 3D $x$ -space learning module based on the transformer. Extensive experiments on data from the human connectome project show that our 3D-HGT outperforms state-of-the-art methods, including HGT, in both quantitative and qualitative evaluations.

Functional connectivity (FC) "gradients" enable investigation of connection topography in relation to cognitive hierarchy, and yield the primary axes along which FC is organized. In this work, we employ a variant of the "gradient" approach wherein we solve for the normal modes of FC, yielding functional connectome harmonics. Until now, research in this vein has only considered static FC, neglecting the possibility that the principal axes of FC may depend on the timescale at which they are computed. Recent work suggests that momentary activation patterns, or brain states, mediate the dominant components of functional connectivity, suggesting that the principal axes may be invariant to change in timescale. In light of this, we compute functional connectome harmonics using time windows of varying lengths and demonstrate that they are stable across timescales. Our connectome harmonics correspond to meaningful brain states. The activation strength of the brain states, as well as their inter-relationships, are found to be reproducible for individuals. Further, we utilize our time-varying functional connectome harmonics to formulate a simple and elegant method for computing cortical flexibility at vertex resolution and demonstrate qualitative similarity between flexibility maps from our method and a method standard in the literature.