Proceedings. IEEE International Symposium on Biomedical Imaging最新文献

The off-the-shelf model for unsupervised domain adaptation (OSUDA) has been introduced to protect patient data privacy and intellectual property of the source domain without access to the labeled source domain data. Yet, an off-the-shelf diagnosis model, deliberately compromised by backdoor attacks during the source domain training phase, can function as a parasite-host, disseminating the backdoor to the target domain model during the OSUDA stage. Because of limitations in accessing or controlling the source domain training data, OSUDA can make the target domain model highly vulnerable and susceptible to prominent attacks. To sidestep this issue, we propose to quantify the channel-wise backdoor sensitivity via a Lipschitz constant and, explicitly, eliminate the backdoor infection by overwriting the backdoor-related channel kernels with random initialization. Furthermore, we propose to employ an auxiliary model with a full source model to ensure accurate pseudo-labeling, taking into account the controllable, clean target training data in OSUDA. We validate our framework using a multi-center, multi-vendor, and multi-disease (M&M) cardiac dataset. Our findings suggest that the target model is susceptible to backdoor attacks during OSUDA, and our defense mechanism effectively mitigates the infection of target domain victims.

Diffusion MRI (dMRI) has shown significant promise in capturing subtle changes in neural microstructure caused by neurodegenerative disorders. In this paper, we propose a novel end-to-end compound architecture for processing raw dMRI data. It consists of a 3D convolutional kernel network (CKN) that extracts macro-architectural features across voxels and a gauge equivariant Volterra network (GEVNet) on the sphere that extracts micro-architectural features from within voxels. The use of higher order convolutions enables our architecture to model spatially extended nonlinear interactions across the applied diffusion-sensitizing magnetic field gradients. The compound network is globally equivariant to 3D translations and locally equivariant to 3D rotations. We demonstrate the efficacy of our model on the classification of neurodegenerative disorders.

In medical image segmentation, although multi-modality training is possible, clinical translation is challenged by the limited availability of all image types for a given patient. Different from typical segmentation models, modality-agnostic (MAG) learning trains a single model based on all available modalities but remains input-agnostic, allowing a single model to produce accurate segmentation given any modality combinations. In this paper, we propose a novel frame-work, MAG learning through Multi-modality Self-distillation (MAG-MS), for medical image segmentation. MAG-MS distills knowledge from the fusion of multiple modalities and applies it to enhance representation learning for individual modalities. This makes it an adaptable and efficient solution for handling limited modalities during testing scenarios. Our extensive experiments on benchmark datasets demonstrate its superior segmentation accuracy, MAG robustness, and efficiency than the current state-of-the-art methods.

Convolutional neural networks (CNNs) have been widely used to build deep learning models for medical image registration, but manually designed network architectures are not necessarily optimal. This paper presents a hierarchical NAS framework (HNAS-Reg), consisting of both convolutional operation search and network topology search, to identify the optimal network architecture for deformable medical image registration. To mitigate the computational overhead and memory constraints, a partial channel strategy is utilized without losing optimization quality. Experiments on three datasets, consisting of 636 T1-weighted magnetic resonance images (MRIs), have demonstrated that the proposal method can build a deep learning model with improved image registration accuracy and reduced model size, compared with state-of-the-art image registration approaches, including one representative traditional approach and two unsupervised learning-based approaches.

Curriculum learning is a learning method that trains models in a meaningful order from easier to harder samples. A key here is to devise automatic and objective difficulty measures of samples. In the medical domain, previous work applied domain knowledge from human experts to qualitatively assess classification difficulty of medical images to guide curriculum learning, which requires extra annotation efforts, relies on subjective human experience, and may introduce bias. In this work, we propose a new automated curriculum learning technique using the variance of gradients (VoG) to compute an objective difficulty measure of samples and evaluated its effects on elbow fracture classification from X-ray images. Specifically, we used VoG as a metric to rank each sample in terms of the classification difficulty, where high VoG scores indicate more difficult cases for classification, to guide the curriculum training process We compared the proposed technique to a baseline (without curriculum learning), a previous method that used human annotations on classification difficulty, and anti-curriculum learning. Our experiment results showed comparable and higher performance for the binary and multi-class bone fracture classification tasks.

We develop deep clustering survival machines to simultaneously predict survival information and characterize data heterogeneity that is not typically modeled by conventional survival analysis methods. By modeling timing information of survival data generatively with a mixture of parametric distributions, referred to as expert distributions, our method learns weights of the expert distributions for individual instances based on their features discriminatively such that each instance's survival information can be characterized by a weighted combination of the learned expert distributions. Extensive experiments on both real and synthetic datasets have demonstrated that our method is capable of obtaining promising clustering results and competitive time-to-event predicting performance.

Idiopathic Rem sleep Behavior Disorder (iRBD) is a significant biomarker for the development of alpha-synucleinopathies, such as Parkinson's disease (PD) or Dementia with Lewy bodies (DLB). Methods to identify patterns in iRBD patients can help in the prediction of the future conversion to these diseases during the long prodromal phase when symptoms are non-specific. These methods are essential for disease management and clinical trial recruitment. Brain PET scans with 18F-FDG PET radiotracers have recently shown promise, however, the scarcity of longitudinal data and PD/DLB conversion information makes the use of representation learning approaches such as deep convolutional networks not feasible if trained in a supervised manner. In this work, we propose a self-supervised learning strategy to learn features by comparing the brain hemispheres of iRBD non-convertor subjects, which allows for pre-training a convolutional network on a small data regimen. We introduce a loss function called hemisphere dissimilarity loss (HDL), which extends the Barlow Twins loss, that promotes the creation of invariant and non-redundant features for brain hemispheres of the same subject, and the opposite for hemispheres of different subjects. This loss enables the pre-training of a network without any information about the disease, which is then used to generate full brain feature vectors that are fine-tuned to two downstream tasks: follow-up conversion, and the type of conversion (PD or DLB) using baseline 18F-FDG PET. In our results, we find that the HDL outperforms the variational autoencoder with different forms of inputs.

In population and longitudinal imaging studies that employ deformable image registration, more accurate results can be achieved by initializing deformable registration with the results of affine registration where global misalignments have been considerably reduced. Such affine registration, however, is limited to linear transformations and it cannot account for large nonlinear anatomical variations, such as those between pre- and post-operative images or across different subject anatomies. In this work, we introduce a new intermediate deformable image registration (IDIR) technique that recovers large deformations via windowed cross-correlation, and provide an efficient implementation based on the fast Fourier transform. We evaluate our method on 2D X-ray and 3D magnetic resonance images, demonstrating its ability to align substantial nonlinear anatomical variations within a few iterations.

The foveal avascular zone (FAZ) is a retinal area devoid of capillaries and associated with multiple retinal pathologies and visual acuity. Optical Coherence Tomography Angiography (OCT-A) is a very effective means of visualizing retinal vascular and avascular areas, but its use remains limited to research settings due to its complex optics limiting availability. On the other hand, fundus photography is widely available and often adopted in population studies. In this work, we test the feasibility of estimating the FAZ from fundus photos using three different approaches. The first two approaches rely on pixel-level and image-level FAZ information to segment FAZ pixels and regress FAZ area, respectively. The third is a training mask-free pipeline combining saliency maps with an active contours approach to segment FAZ pixels while being trained on image-level measures of the FAZ areas. This enables training FAZ segmentation methods without manual alignment of fundus and OCT-A images, a time-consuming process, which limits the dataset that can be used for training. Segmentation methods trained on pixel-level labels and image-level labels had good agreement with masks from a human grader (respectively DICE of 0.45 and 0.4). Results indicate the feasibility of using fundus images as a proxy to estimate the FAZ when angiography data is not available.

In order to quantify lateral asymmetric degeneration of the hippocampus for early predicting Alzheimer's disease (AD), we develop a deep learning (DL) model to learn informative features from the hippocampal magnetic resonance imaging (MRI) data for predicting AD conversion in a time-to-event prediction modeling framework. The DL model is trained on unilateral hippocampal data with an autoencoder based regularizer, facilitating quantification of lateral asymmetry in the hippocampal prediction power of AD conversion and identification of the optimal strategy to integrate the bilateral hippocampal MRI data for predicting AD. Experimental results on MRI scans of 1307 subjects (817 for training and 490 for validation) have demonstrated that the left hippocampus can better predict AD than the right hippocampus, and an integration of the bilateral hippocampal data with the instance based DL method improved AD prediction, compared with alternative predictive modeling strategies.