Diabetic retinopathy (DR) and diabetic macular edema (DME) are the leading causes of permanent blindness in working-age population. Automatic grading DR DME helps ophthalmologists design tailored treatments to patients, thus is vital importance clinical practice. However, prior works either grade or DME, ignore correlation between its complication, i.e., DME. Moreover, location information, e.g., macula soft hard exhaust annotations, widely used as a for grading. Such annotations costly...

Rain is a common weather phenomenon, where object visibility varies with depth from the camera and objects faraway are visually blocked more by fog than rain streaks. Existing methods datasets for removal, however, ignore these physical properties, thereby limiting removal efficiency on real photos. In this work, we first analyze visual effects of subject to scene formulate imaging model collectively streaks fog; then, prepare new dataset called RainCityscapes outdoor Furthermore, design an...

We present a novel method for removing rain streaks from single input image by decomposing it into rain-free background layer B and rain-streak R. A joint optimization process is used that alternates between details non-streak The assisted three priors. Observing typically span narrow range of directions, we first analyze the local gradient statistics in to identify regions are dominated streaks. From these regions, estimate dominant streak direction extract collection rain-dominated...

Shadow detection and shadow removal are fundamental challenging tasks, requiring an understanding of the global image semantics. This paper presents a novel deep neural network design for by analyzing spatial context in direction-aware manner. To achieve this, we first formulate attention mechanism recurrent (RNN) introducing weights when aggregating features RNN. By learning these through training, can recover (DSC) detecting removing shadows. is developed into DSC module embedded...

Automatic prostate segmentation in transrectal ultrasound (TRUS) images is of essential importance for image-guided interventions and treatment planning. However, developing such automatic solutions remains very challenging due to the missing/ambiguous boundary inhomogeneous intensity distribution TRUS, as well large variability shapes. This paper develops a novel 3D deep neural network equipped with attention modules better TRUS by fully exploiting complementary information encoded...

Recently neural architecture (NAS) search has attracted great interest in academia and industry. It remains a challenging problem due to the huge space computational costs. Recent studies NAS mainly focused on usage of weight sharing train SuperNet once. However, corresponding branch each subnetwork is not guaranteed be fully trained. may only incur computation costs but also affect ranking retraining procedure. We propose multi-teacher-guided NAS, which proposes use adaptive ensemble...

Masked image modeling (MIM) with transformer backbones has recently been exploited as a powerful self-supervised pre-training technique. The existing MIM methods adopt the strategy to mask random patches of and reconstruct missing pixels, which only considers semantic information at lower level, causes long time. This paper presents HybridMIM, novel hybrid learning method based on masked for 3D medical segmentation. Specifically, we design two-level masking hierarchy specify how in...

This paper presents a deep multi-model fusion network to attentively integrate multiple models separate layers and boost the performance in single-image dehazing. To do so, we first formulate attentional feature integration module maximize of convolutional neural (CNN) features at different CNN generate multi-level integrated (AMLIF). Then, from AMLIF, further predict haze-free result for an atmospheric scattering model, as well four haze-layer separation models, then fuse results together...

We address the problem of correcting exposure underexposed photos. Previous methods have tackled this from many different perspectives and achieved remarkable progress. However, they usually fail to produce natural-looking results due existence visual artifacts such as color distortion, loss detail, inconsistency, etc. find that main reason why existing induce these is because break a perceptually similarity between input output. Based on observation, an effective criterion, termed...

Single image dehazing is a challenging task, for which the domain shift between synthetic training data and real-world testing images usually leads to degradation of existing methods. To address this issue, we propose novel framework collaborating with unlabeled real data. First, develop disentangled network (DID-Net), disentangles feature representations into three component maps, i.e. latent haze-free image, transmission map, global atmospheric light estimate, respecting physical model...

This paper presents a new deep neural network design for salient object detection by maximizing the integration of local and global image context within, around, beyond objects.Our key idea is to adaptively propagate aggregate features with variable attenuation over entire feature maps.To achieve this, we spatial (SAC) module recurrently translate independently different factors then attentively learn weights integrate aggregated features.By further embedding process individual layers in...

Specular highlight detection and removal are fundamental challenging tasks. Although recent methods have achieved promising results on the two tasks by training synthetic data in a supervised manner, they typically solely designed for or removal, their performance usually deteriorates significantly real-world images. In this paper, we present novel network that aims to detect remove highlights from natural To domain gap between samples real test images, support investigation of...

Camouflaged object detection is a challenging task that aims to identify objects having similar texture the surroundings. This paper presents amplify subtle difference between camouflaged and background for by formulating multiple texture-aware refinement modules learn features in deep convolutional neural network. The module computes biased co-variance matrices of feature responses extract information, adopts an affinity loss set parameter maps help separate background, leverages...

Rain is a common weather phenomenon that affects environmental monitoring and surveillance systems. According to an established rain model (Garg Nayar, 2007), the scene visibility in varies with depth from camera, where objects faraway are visually blocked more by fog than streaks. However, existing datasets methods for removal ignore these physical properties, thus limiting efficiency on real photos. In this work, we analyze visual effects of subject formulate imaging collectively considers...

Pancreatic cancer has the worst prognosis of all cancers. The clinical application endoscopic ultrasound (EUS) for assessment pancreatic risk and deep learning classification EUS images have been hindered by inter-grader variability labeling capability. One key reasons these difficulties is that are obtained from multiple sources with varying resolutions, effective regions, interference signals, making distribution data highly variable negatively impacting performance models. Additionally,...

Surgical instrument segmentation is fundamentally important for facilitating cognitive intelligence in robot-assisted surgery. Although existing methods have achieved accurate results, they simultaneously generate masks of all instruments, which lack the capability to specify a target object and allow an interactive experience. This paper focuses on novel essential task robotic surgery, i.e., Referring Video Instrument Segmentation (RSVIS), aims automatically identify segment surgical...

Recently, Denoising Diffusion Models have achieved outstanding success in generative image modeling and attracted significant attention the computer vision community. Although a substantial amount of diffusion-based research has focused on tasks, few studies apply diffusion models to medical diagnosis. In this paper, we propose network (named DiffMIC-v2) address general classification by eliminating unexpected noise perturbations representations. To achieve goal, first devise an improved...

Speckle refers to the granular patterns that occur in ultrasound images due wave interference. removal can greatly improve visibility of underlying structures an image and enhance subsequent post processing. We present a novel framework for speckle based on low-rank non-local filtering. Our approach works by first computing guidance assists selection candidate patches filtering face significant speckles. The are further refined using minimization estimated truncated weighted nuclear norm...

This paper presents a novel deep learning model to aggregate the attentional dilated features for salient object detection by exploring complementary information between global and local context in convolutional neural network. There are two technical contributions our network design. First, we develop an dense atrous (dilated) spatial pyramid pooling (AD-ASPP) module selectively use saliency cues captured convolutions with small rate large rate. Second, taking feature as backbone,...

This article presents a deep normal filtering network, called DNF-Net, for mesh denoising. To better capture local geometry, our network processes the in terms of patches extracted from mesh. Overall, DNF-Net is an end-to-end that takes facet normals as inputs and directly outputs corresponding denoised patches. In this way, we can reconstruct geometry with feature preservation. Besides overall architecture, contributions include novel multi-scale embedding unit, residual learning strategy...

This work presents a gated non-local deep residual learning framework for image deraining. It can avoid the over-deraining or under-deraining caused by global in existing deraining networks, since learned soft gate our method adaptively adjusts amount of to be passed generating final derained result. To generate feature maps prediction, we develop guided attention module (NLAM), which first obtains features exploiting spatial inter-dependencies among all positions local produced...