Depression remains a pervasive mental health disorder that demands prompt diagnosis and intervention. Although social media data presents a promising avenue for early detection, traditional deep neural models are frequently critiqued for their lack of interpretability and susceptibility to bias. We introduce ProtoDep—a neurosymbolic framework that integrates clinically grounded categorizations (e.g., PHQ-9 symptoms) with large language model–assisted prototype learning. Unlike conventional black-box models, ProtoDep aligns individual tweets with symptom-level prototypes, offering interpretable explanations at three levels: (i) symptom-level insights that map user posts to recognized depressive patterns, (ii) case-based reasoning that compares users to representative prototype profiles, and (iii) transparent concept-level decisions, wherein classification at inference time is driven by the distances between the user profile and prototype user and symptom clusters, yielding clear, quantifiable explanations. By integrating symbolic mental health constructs with neural embeddings, ProtoDep achieves a mean F1-score of 94% across five benchmark datasets and establishes a foundation for interpretable depression screening pipelines with potential applicability in clinical settings.

In this paper, a multimodal deep architecture for classification of light detection and ranging (LiDAR) and hyperspectral image (HSI) is proposed, acquiring the knowledge of both modalities by leveraging modality specific information as well as their complementary information. The proposed model consists of two main steps. First, to improve the performance of a two-dimensional convolutional neural network (2DCNN), low-frequency features with maximum autocorrelation factor of HSI are injected into 2DCNN which are called multi-scale features of 2DCNN. Second, to improve the accuracy of 2DCNN and extract smooth and semantic information, the posterior energy of hidden Markov random field (HMRF) is modified by using Gaussian attention and albedo recovery attention mechanisms and energies of LiDAR and HMRF. Then, these features are fused based on another attention mechanism called attention-based HMRF. Moreover, this HMRF model is used for fusion of HSI and LiDAR. The proposed model is tested on the Houston 2013, Trento and MUUFL datasets and compared with several state-of-the-art methods. The resulting classification accuracies through ablation study show the superior performance of the proposed method.

Multimodal data provide complementary information of a natural phenomenon by integrating data from various domains with very different statistical properties. Capturing the intramodality and cross-modality information of multimodal data is the essential capability of multimodal learning methods. The geometry-aware data analysis approaches provide these capabilities by implicitly representing data in various modalities based on their geometric underlying structures. Also, in many applications, data are explicitly defined on an intrinsic geometric structure. Generalizing deep learning methods to the non-Euclidean domains is an emerging research field, which has recently been investigated in many studies. Most of those popular methods are developed for unimodal data. In this article, a multimodal graph wavelet convolutional network (M-GWCN) is proposed as an end-to-end network. M-GWCN simultaneously finds intramodality representation by applying the multiscale graph wavelet transform to provide helpful localization properties in the graph domain of each modality and cross-modality representation by learning permutations that encode correlations among various modalities. M-GWCN is not limited to either the homogeneous modalities with the same number of data or any prior knowledge indicating correspondences between modalities. Several semisupervised node classification experiments have been conducted on three popular unimodal explicit graph-based datasets and five multimodal implicit ones. The experimental results indicate the superiority and effectiveness of the proposed methods compared with both spectral graph domain convolutional neural networks and state-of-the-art multimodal methods.

Abstract Due to the American cancer society, many people with esophageal adenocarcinoma are not survived. The treatment rate can be significant in the early detection of Barrett's esophagus (BE) as a premalignant stage for adenocarcinoma. An important landmark to detect BE is the Z‐line. BE segmentation is already highly dependent upon the operator's knowledge and skill. The main aim of this study is automatic Z‐line extraction using endoscopic images leading to segmentation of the early BE stage. To this end, a computer‐aided detection method exploiting k‐means clustering, image segmentation using the edge detector, and a novel boundary linking algorithm is proposed. For the evaluation, the gold standard is considered the average contours of Z‐lines extracted by the three experts. The proposed method annotated the Z‐line with the accuracy and precision of 0.92 and 0.87, respectively, and the value of the average boundary distance is 5.9 pixels. To the results and visual inspection, the presented method can be used for efficient and robust extraction of the Z‐line at the early BE stage. Furthermore, it can be used in other medical imaging applications with complex boundaries and low contrast in the images, limiting the common automatic boundary detection methods.

Gastroesophageal reflux disease (GERD) is a common digestive disorder with troublesome symptoms that has been affected millions of people worldwide. Multichannel Intraluminal Impedance-pH (MII-pH) monitoring is a recently developed technique, which is currently considered as the gold standard for the diagnosis of GERD. In this paper, we address the problem of characterizing gastroesophageal reflux events in MII signals. A GER detection algorithm has been developed based on the sparse representation of local segments. Two dictionaries are trained using the online dictionary learning approach from the distal impedance data of selected patches of GER and no specific patterns intervals. A classifier is then designed based on thelp-norm of dictionary approximations. Next, a preliminary permutation mask is obtained from the classification results of patches, which is then used in post-processing procedure to investigate the exact timings of GERs at all impedance sites. Our algorithm was tested on 33 MII episodes, resulting a sensitivity of 96.97% and a positive predictive value of 94.12%.

Localisation of an active capsule endoscope inside the stomach has different challenges. One of them is the estimation of the capsule's roll angle. Another challenge is adjusting the distance between the capsule and the stomach to achieve high‐quality imaging in the region of interest. In this study, an optimised image‐guided localisation (O‐Localisation) method is proposed to estimate the roll angle and the scale factor between the consecutive frames. The distance between the capsule and walls of the stomach can be adjusted using the suggested fuzzy adjuster, which is developed based on the estimated scale factors and calibration parameters. This new method is only based on visual information extracted from wireless capsule endoscope video frames. The results show that this method can accurately estimate the rotation angles and scale factors with errors <0.2% for the angles up to 90° and 0.3% for the scales up to 5, respectively. The method is robust to the brightness changes up to 80% with a maximum error of 0.3%. The computational time is about 1 s and can be considered near real‐time for this application. Accordingly, the O‐Localisation method as a real‐time, robust and precise method for capsule localisation can provide a more efficient controllable and steerable capsule endoscopes.

Hierarchical graphical models can incorporate jointly several tasks in a unified framework. By applying this approach, information exchange among tasks would improve the results. A hierarchical conditional random field (CRF) is proposed here to improve the semantic image segmentation. Although this newly proposed model applies the information of several tasks, its run time is comparable with the contemporary approaches. This method is evaluated on MSRC dataset and has shown similar or better segmentation accuracy in comparison with models where CRFs or hierarchical models are adopted.

In this study, first a supervised version for probabilistic principal component analysis mixture model is proposed. Using this model, local linear underlying manifolds of data samples are obtained. These underlying manifolds are used in a dimensionality reduction without loss framework, for face recognition application. In this framework, the benefits of dimensionality reduction are used in the predictive model, while using the projection penalty idea, the loss of useful information will be minimised. The authors use support vector machine (SVM) and k ‐nearest neighbour (KNN) classifiers as the predictive models in this framework. To train and evaluate the proposed method, the well‐known face databases are used. The experimental results show that the proposed method with SVM as the predictive model have the most average classification accuracy compared with many traditional methods which use predictive model SVM after dimensionality reduction, and also compared with the projection penalty idea used for linear and non‐linear kernel‐based dimensionality reduction methods. Moreover, their experiments show that the proposed method with KNN as predictive model is superior to the case that dimensionality reduction is performed, and then the KNN classifier is applied.

This paper introduces an unsupervised batch algorithm for learning the underlying regional linear manifolds and estimating their dimensionalities using a data set in a topographic map. For this purpose, a unified free energy functional is designed and an expectation-maximization procedure is developed to minimize it. Regional dimensionality estimation controls the extent of the linear manifolds. This property makes the model appropriate for representing the datasets with varying regional intrinsic dimensions, compared to the resembling techniques without dimensionality learning capability. Experimental results show the good performance of the model on synthesized and realworld applications.

This paper introduces the kernel-based linear manifold topographic map and an information theoretic algorithm developed for its learning. The kernels represent lower dimensional local linear manifolds in a data space, and are defined in an optimal manner when special Gaussian input densities are assumed. The kernel parameters are adapted to maximize the joint entropy of the neuron outputs of the map. This is fulfilled by applying stochastic gradient ascent to the differential entropy of each neuron output and using competition between the neurons of the map. Topology preserving property is also possible by considering neighborhood functions. The proposed model can be considered as an improved version of the ASSOM network which maintains the ASSOM advantages while avoiding its limitations.