This paper proposes a novel variational inference framework, SCVAE, for cardiac image segmentation. SCVAE embeds a manifold space between the image space and the image representation space using a VAE, while the representation space is learned through contrastive learning. The manifold not only captures semantic feature categories but also enables the reconstruction of the image representation space, thereby providing a regularization mechanism for learning image representations. We theoretically analyze the relationship among the input space, manifold space, and representation space, and prove that, under optimal parameterization, the KL divergence between the variational posterior and the true posterior over the dataset achieves the smallest. Moreover, an approximated optimal prior is constructed in our SCVAE, and the prior and posterior distributions provide a more reliable uncertainty map to guide sample selection in contrastive learning. Our framework is general and can be embedded into various contrastive learning networks for segmentation tasks, enabling a mutually beneficial coupling of generative modeling and discriminative feature learning. Extensive experiments on benchmark cardiac image datasets validate the superior performance of SCVAE compared to state-of-the-art semi-supervised segmentation models.

Recovering spatial-varying bi-directional reflectance distribution function (SVBRDF) from as few as possible captured images has been a challenging task in computer graphics. Benefiting from the co-located flashlight-camera capture strategy and data-driven priors, SVBRDF can be estimated from few input images. However, this capture strategy usually requires a controllable darkroom environment, ensuring the flashlight is a single light source. It is often impractical during on-site capture in real-world scenarios. To support SVBRDF estimation in an uncontrolled environment, the key challenge lies in the high-precise estimation of unknown environment lighting and its effective utilization on SVBRDF recovery. To address this issue, we proposed a novel exemplar-based environment lighting representation, which is easier to use for neural networks. These exemplars are a set of rendered images of selected materials under the environment lighting. By embedding the rendering process, our approach transforms environment lighting represented in the spherical domain into the sample-surface domain, thereby achieving the domain alignment with input images. This significantly reduces the network’s learning burden, resulting in a more precise environment lighting estimation. Furthermore, after lighting prediction, we also present a dominant lighting extraction algorithm and an adaptive exemplar selection algorithm to enhance the guidance of environment lighting in SVBRDF estimation. Finally, considering the distant contribution of environment lighting and point lighting to SVBRDF recovery, we proposed a well-designed cascaded network. Quantitative assessments and qualitative analysis have demonstrated that our method achieves superior SVBRDF estimations compared to previous approaches. The source code will be released.

This study introduces a novel generative adversarial network (GAN)-based Dual-stage Teacher-Student Representation Learning (GDL) framework designed to extract effective representations from unlabeled data for cardiac image segmentation. The GDL framework employs an adversarial augmentation approach, simultaneously training a generation model to produce diverse samples and a teacher-student model to learn image representations iteratively. During the pre-training stage, a generator is trained using a diversity-enhanced contrastive loss to generate diverse augmented samples. Subsequently, in the first stage, the generator is further supervised by enforcing dissimilarity between representations produced by the teacher and student models, aiming to expand the image representation space of generated images. In the second stage, the student model is trained to transfer distilled knowledge acquired from the augmented samples to the teacher model to improve its generalization performance in the image segmentation task. Our proposed model ensures that the diverse generated samples correspond precisely to their pseudo-labels, facilitating learning semantic representations from unlabeled data. Experimental results on public cardiac MR datasets demonstrate that our model surpasses the performance of state-of-the-art semi-supervised learning networks.

The dendritic neural model (DNM) is computationally faster than other machine-learning techniques, because its architecture can be implemented by using logic circuits and its calculations can be performed entirely in binary form. To further improve the computational speed, a straightforward approach is to generate a more concise architecture for the DNM. Actually, the architecture search is a large-scale multiobjective optimization problem (LSMOP), where a large number of parameters need to be set with the aim of optimizing accuracy and structural complexity simultaneously. However, the issues of irregular Pareto front, objective discontinuity, and population degeneration strongly limit the performances of conventional multiobjective evolutionary algorithms (MOEAs) on the specific problem. Therefore, a novel competitive decomposition-based MOEA is proposed in this study, which decomposes the original problem into several constrained subproblems, with neighboring subproblems sharing overlapping regions in the objective space. The solutions in the overlapping regions participate in environmental selection for the neighboring subproblems and then propagate the selection pressure throughout the entire population. Experimental results demonstrate that the proposed algorithm can possess a more powerful optimization ability than the state-of-the-art MOEAs. Furthermore, both the DNM itself and its hardware implementation can achieve very competitive classification performances when trained by the proposed algorithm, compared with numerous widely used machine-learning approaches.