MRNet's feature extraction process is composed of concurrent convolutional and permutator-based pathways, utilizing a mutual information transfer module to harmonize feature exchanges and correct inherent spatial perception biases for better representation quality. To mitigate the bias introduced by pseudo-label selection, RFC dynamically adjusts the strong and weak augmented distributions to ensure a rational discrepancy, and augments features for underrepresented categories to establish balanced training. During momentum optimization, the CMH model, in an effort to counteract confirmation bias, mirrors the consistency of different sample augmentations within the network's update process, consequently strengthening the model's dependability. Extensive research conducted on three semi-supervised medical image categorization datasets showcases HABIT's efficacy in diminishing three biases, achieving groundbreaking results. Our project's code repository is located at https://github.com/CityU-AIM-Group/HABIT.
Recent advancements in vision transformers have sparked a surge of interest in medical image analysis, thanks to their exceptional performance across numerous computer vision applications. While recent hybrid/transformer-based approaches prioritize the strengths of transformers in capturing long-distance dependencies, they often fail to acknowledge the issues of their significant computational complexity, substantial training costs, and superfluous interdependencies. This paper introduces an adaptive pruning technique for transformer-based medical image segmentation, resulting in the lightweight and effective APFormer hybrid network. nonalcoholic steatohepatitis Our investigation reveals that this is the first instance of transformer pruning used for medical image analysis tasks. APFormer's key strengths lie in its self-regularized self-attention (SSA), which improves the convergence of dependency establishment, its Gaussian-prior relative position embedding (GRPE), which enhances the learning of positional information, and its adaptive pruning, which minimizes redundant calculations and perceptual input. The well-converged dependency distribution and Gaussian heatmap distribution, employed by SSA and GRPE, serve as prior knowledge for self-attention and position embeddings, respectively, facilitating transformer training and providing a solid basis for the pruning steps that follow. electron mediators For both query-wise and dependency-wise pruning, adaptive transformer pruning modifies gate control parameters to achieve performance improvement and complexity reduction. The substantial segmentation performance of APFormer, against state-of-the-art models, is confirmed by exhaustive experiments on two frequently utilized datasets, accompanied by a lower parameter count and lower GFLOPs. Importantly, empirical ablation studies validate that adaptive pruning can be integrated effortlessly, boosting performance across diverse hybrid and transformer-based architectures. The APFormer project's code is downloadable from https://github.com/xianlin7/APFormer.
The precise delivery of radiotherapy, a hallmark of adaptive radiation therapy (ART), requires the careful adaptation to anatomical changes. The synthesis of computed tomography (CT) from cone-beam CT (CBCT) is an essential part of this process. Consequently, the issue of substantial motion artifacts makes CBCT-to-CT synthesis in breast-cancer ART applications a significant challenge. The omission of motion artifacts from existing synthesis methods compromises their performance in chest CBCT image analysis. This paper approaches CBCT-to-CT synthesis by dividing it into the two parts of artifact reduction and intensity correction, aided by breath-hold CBCT image data. Our multimodal unsupervised representation disentanglement (MURD) learning framework, designed to achieve superior synthesis performance, disentangles the content, style, and artifact representations of CBCT and CT images within the latent space. Image variety is produced by MURD through the recombination of its disentangled image representations. Our approach integrates a multipath consistency loss for improved structural consistency in synthesis, and a multi-domain generator to amplify synthesis performance. MURD, evaluated on our breast-cancer dataset, exhibited striking performance in synthetic CT, with a mean absolute error of 5523994 HU, a structural similarity index of 0.7210042, and a peak signal-to-noise ratio of 2826193 dB. Compared to cutting-edge unsupervised synthesis techniques, our approach yields enhanced synthetic CT images, demonstrating improvements in both accuracy and visual appeal within the results.
For unsupervised domain adaptation in image segmentation, we describe a method that aligns high-order statistics from source and target domains to detect domain-invariant spatial relationships among segmentation categories. Our method's initial step involves estimating the joint probability distribution of predictions for pixel pairs exhibiting a predetermined spatial relationship. Domain adaptation is subsequently accomplished by aligning the combined probability distributions of source and target images, determined for a collection of displacements. Enhancing this process in two ways is recommended. A multi-scale strategy, highly effective, captures long-range statistical relationships. The joint distribution alignment loss, in the second approach, is extended to encompass features within the network's intermediate layers, a process achieved via cross-correlation computation. We examine our method's performance on the task of unpaired multi-modal cardiac segmentation, particularly on the Multi-Modality Whole Heart Segmentation Challenge dataset, as well as the prostate segmentation task, incorporating images from two datasets that are taken from different domains. check details Our research demonstrates the advantages of our approach when evaluating it against current methods for cross-domain image segmentation. The Domain adaptation shape prior's project files are located on the Github page at https//github.com/WangPing521/Domain adaptation shape prior.
We describe a novel video-based method, employing non-contact technology, for recognizing elevated skin temperatures exceeding typical ranges in individuals. A critical diagnostic step involves recognizing elevated skin temperatures, which can signal infection or a medical problem. Typically, contact thermometers or non-contact infrared-based sensors are utilized to detect elevated skin temperatures. Given the widespread use of video data acquisition devices like mobile phones and personal computers, a binary classification system, Video-based TEMPerature (V-TEMP), is constructed to categorize subjects displaying either normal or elevated skin temperatures. The empirical distinction between skin at normal and elevated temperatures is achieved through exploiting the correlation between skin temperature and the angular reflectance of light. This correlation's uniqueness is illustrated by 1) revealing a difference in the angular distribution of light reflected from skin-like and non-skin-like materials and 2) exploring the uniformity in the angular distribution of light reflected from materials with optical properties akin to human skin. Ultimately, we showcase V-TEMP's resilience by assessing the effectiveness of elevated skin temperature identification on subject recordings acquired in 1) controlled laboratory settings and 2) real-world, outdoor scenarios. V-TEMP's positive attributes include: (1) the elimination of physical contact, thus reducing the potential for infections transmitted via physical interaction, and (2) the capacity for scalability, which leverages the prevalence of video recording devices.
Portable tools for monitoring and identifying daily activities have become a growing focus in digital healthcare, particularly for the elderly. A substantial problem in this domain arises from the considerable dependence on labeled activity data for effectively developing corresponding recognition models. Labeled activity data is a resource demanding considerable expense to collect. To overcome this predicament, we propose a strong and dependable semi-supervised active learning technique, CASL, which amalgamates prevalent semi-supervised learning strategies with a mechanism for expert collaboration. Input to CASL is exclusively the user's trajectory. In conjunction with expert collaboration, CASL analyzes valuable model samples for optimized performance. CASL's exceptional activity recognition performance stems from its minimal reliance on semantic activities, outpacing all baseline methods and achieving a level of performance similar to that of supervised learning methods. CASL exhibited 89.07% accuracy on the adlnormal dataset, featuring 200 semantic activities, in comparison to supervised learning's superior 91.77% accuracy. In our CASL, a query strategy and a data fusion approach were essential in the validation process performed by the ablation study of the components.
In the world, Parkinson's disease commonly afflicts the middle-aged and elderly demographic. Currently, clinical assessment forms the cornerstone of Parkinson's disease diagnosis, yet diagnostic accuracy remains suboptimal, particularly in the initial stages of the illness. Employing a deep learning hyperparameter optimization approach, this paper proposes a novel Parkinson's auxiliary diagnostic algorithm for the identification of Parkinson's disease. ResNet50, employed by the diagnostic system for feature extraction and Parkinson's classification, encompasses speech signal processing, Artificial Bee Colony (ABC) algorithm-based enhancements, and ResNet50 hyperparameter optimization. The GDABC algorithm, an improved Artificial Bee Colony algorithm, incorporates a Range pruning strategy, to constrain the search area, and a Dimension adjustment strategy, to modify the gbest dimension individually for each dimension. The verification set of the Mobile Device Voice Recordings (MDVR-CKL) dataset, collected at King's College London, exhibits a diagnosis system accuracy greater than 96%. Our auxiliary diagnostic system for Parkinson's disease demonstrates superior classification performance on the dataset when benchmarked against current sound-based diagnostic approaches and optimized algorithms, given the constraints of available time and resources.