Conducting survival analysis on Intensive Care Unit (ICU) data enables us to understand how various factors affect patients’ survival time, guiding clinical decisions and treatment plans. Current ICU analysis algorithms generally lack robustness and potentially lead to erroneous and misleading outcomes as they struggle to manually identify small yet effective subsets within large long-tail clinical datasets. Traditionally, the Synthetic Minority Oversampling Technique (SMOTE) is simple and powerful for tabular data but has difficulty generating discrete features. Furthermore, current generative-based balancing methods often perform poorly compared to SMOTE and easily experience mode collapse, which is crucial for ICU analysis. To address these issues, this study first introduces a novel cohort prioritization technique to shortlist effectively relevant cohorts. Cohort Prioritization is a feature-selective approach that identifies pertinent cohorts within the dataset based on the proposed index score system. To tackle the imbalance issue, we proposed an innovative Conditional Generative Adversarial Network (cGAN), cGAN-based SMOTE improved approach called Intensive Care Data Balancer (ICD-Balancer), leveraging the power of generative adversarial networks to rebalance class distributions, where we employed Gumbel-Softmax and proposed Gumbel-Sigmoid to leverage generating multimodal tabular features thereby enhancing the overall performance of classification tasks. The comparative results demonstrate the effectiveness of our proposed approach in resolving class imbalance and the utility of our index scoring system in enhancing feature selection processes. This research advances the state-of-the-art in optimizing feature selection and addressing the class imbalance, with potential applications across diverse tabular medical data requiring robust and interpretable data analysis.

The adoption of artificial intelligence (AI) in healthcare requires not only accurate predictions but also a clear understanding of its reliability. In safety-critical domains such as medical imaging and diagnosis, clinicians must assess the confidence in model outputs to ensure safe decision making. Uncertainty quantification (UQ) addresses this need by providing confidence estimates and identifying situations in which models may fail. Such uncertainty estimates enable risk-aware deployment, improve model robustness, and ultimately strengthen clinical trust. Although prior studies have surveyed UQ in centralized learning, a systematic review in the federated learning (FL) context is still lacking. As a privacy-preserving collaborative paradigm, FL enables institutions to jointly train models without sharing raw patient data. However, compared with centralized learning, FL introduces more complex sources of uncertainty. In addition to data uncertainty caused by noisy inputs and model uncertainty from distributed optimization, there also exists distributional uncertainty arising from client heterogeneity and personalized uncertainty associated with site-specific biases. These intertwined uncertainties complicate model reliability and highlight the urgent need for UQ strategies tailored to federated settings. This survey reviews UQ in medical FL. We categorize uncertainties unique to FL and compare them with those in centralized learning. We examine the sources of uncertainty, existing FL architectures, UQ methods, and their integration with privacy-preserving techniques, and we analyze their advantages, limitations, and trade-offs. Finally, we highlight key challenges—scalable UQ under non-IID conditions, federated OOD detection, and clinical validation—and outline future opportunities such as hybrid UQ strategies and personalization. By combining methodological advances in UQ with application perspectives, this survey provides a structured overview to inform the development of more reliable and privacy-preserving FL systems in healthcare.

Breast cancer remains a leading cause of female mortality globally, creating a significant burden on healthcare systems and underscoring the need for precise AI-driven diagnostic solutions. Whole Slide Images (WSI) are central to breast cancer diagnosis but require substantial storage for high-resolution features, complicating their efficient use in medical applications. Vision Transformers (ViT) offer a promising approach for WSI patch processing, yet traditional ViT-based methods are computationally intensive and overlook cross-sample correlations. To address these limitations, we propose B-EANet, a novel framework that integrates ViT with an external attention mechanism for improved image classification in multi-modal breast cancer diagnosis. B-EANet's architecture includes PatchExtract, which efficiently extracts patches from mammograms, and PatchEmbedding, which reduces intermediate feature maps. The external attention mechanism reduces time complexity to linear, capturing intra-modality relationships while preserving computational efficiency. We evaluated B-EANet on three publicly available datasets, including BreakHis, CAMELYON 2017, and BACH 2018. The proposed approach achieved state-of-the-art results on all three datasets. BreakHis dataset achieved an F1-score of 96.2% and an AUC of 99.9%, while CAMELYON 2017 demonstrated an F1-Score of 93.89%, and AUC of 96.32%, and BACH 2018 recorded an F1-Score of 94.37% and an AUC of 97.71%. The B-EANet utilized only 4.35 MB of memory with 1.1 x 106 parameters. This framework significantly enhances the efficiency and effectiveness of automated breast cancer diagnosis, making deep learning more accessible to medical practice as well as it fulfills demands for resource-constrained clinical scenarios.

PURPOSE:This study investigates the feasibility of utilizing deep learning models to robustly predict patient-specific quality assurance (PSQA) outcomes in fixed field intensity-modulated radiation therapy (FF-IMRT) plans on the Halcyon linear accelerator equipped with a novel dual-layer multi-leaf collimator (MLC). The study explores the integration of Shuffle Attention (SA) mechanisms and deep imbalance regression techniques to enhance the precision and robustness of deep learning-based PSQA predictions. It ensures relative prediction robustness in the extreme imbalance distribution of gamma passing rate (GPR) values. METHODS:Data from 214 FF-IMRT treatment plans covering various treatment sites comprising 1394 beam orientations and corresponding Portal Dosimetry verification data were collected. Fluence maps calculated for each beam orientation served as inputs for the ResNet model. First, the SA module was introduced to improve the prediction accuracy of ResNet, resulting in the proposed Att-ResNet model. Furthermore, to ensure prediction robustness in the GPR values with extreme imbalance distribution, we incorporated the Label Distribution Smoothing (LDS) technique, ultimately forming the ALDS-ResNet method. RESULTS:ALDS-ResNet exhibited smaller mean absolute error (MAE) values than ResNet across all gamma criteria (1%/1 mm: 2.035 vs. 1.824, 2%/2 mm: 1.416 vs. 1.178, 3%/3 mm: 0.951 vs. 0.787). ALDS-ResNet also demonstrated lower MAE values than ResNet for complex but important plan samples (GPR < 85, 1%/1 mm: 10.163 vs. 4.985, 2%/2 mm: 7.443 vs. 3.272, 3%/3 mm: 5.031 vs. 2.940). Compared to ResNet, ALDS-ResNet achieved higher Pearson correlation coefficient (CC) values at 2%/2 mm and 3%/3 mm gamma criteria, measuring 0.7864 and 0.7852, respectively. CONCLUSIONS:The deep learning model based on ResNet shows promise for predicting GPR values in linacs with dual-layer MLC. Integrating attention mechanisms with deep learning networks enhances the accuracy of PSQA predictions. The LDS technique is attributed to the substantial improvement in failed plan GPR prediction accuracy and robustness. Specifically, the deep learning model tailored for dual-layer MLC linacs can be an auxiliary tool for physicists in identifying PSQA failure plans.

Background and objective: In recent times, medical imaging analysis (MIA) has seen an increasing interest due to its core application in computer-aided diagnosis systems (CADs). A modality in MIA refers to a specific technology used to produce human body images, such as MRI, CT scans, or X-rays. Each modality presents unique challenges and characteristics, often leading to imbalances within datasets. This significant challenge impedes model training and generalization due to the varying convergence rates of different modalities and the suppression of gradients in less dominant modalities. Methods: This paper proposes a novel fusion approach, and we named it Slice-Fusion. The proposed approach aims to mitigate the modality imbalance problem by implementing a “Modality-Specific-Balancing-Factor” fusion strategy. Furthermore, it incorporates an auxiliary (uni-modal) task that generates balanced modality pairs based on the image orientations of different modalities. Subsequently, a novel multimodal classification framework is presented to learn from the generated balanced modalities. The effectiveness of the proposed approach is evaluated through comparative assessments on a publicly available BraTS2021 dataset. The results demonstrate the efficiency of Slice-Fusion in resolving the modality imbalance problem. By enhancing the representation of balanced features and reducing modality bias, this approach holds promise for advancing visual health informatics and facilitating more accurate and reliable medical image analysis. Results: In the experiment section, three diverse experiments are conducted such as i) Fusion Loss Metrics Evaluation, ii) Classification, and iii) Visual Health Informatics. Notably, the proposed approach achieved an F1-Score of (100%, 81.25%) on the training and validation sets for the classification generalization task. In addition to the Slice-Fusion’s out-performance, the study also created a new modality-aligned dataset (a highly balanced and informative modality-specific image collection) that aids further research and improves MIA’s robustness. These advancements not only enhance the capability of medical diagnostic tools but also create opportunities for future innovations in the field. Conclusion: This study contributes to advancing medical image analysis, such as effective modality fusion, image reconstruction, comparison, and glioma classification, facilitating more accurate and reliable results, and holds promise for further advancements in visual health informatics.

Background: In computational pathology (CP) analysis, computational efficiency and precise classification outcomes are paramount for robust and scalable solutions. Despite recent advancements in deep-learning frameworks for Whole-Slide Images (WSIs), the heterogeneity of WSIs across different domains poses considerable challenges for developing models with robust generalization capabilities. This study presents the WSI-P2P (Whole-Slide Imaging–Patch to Prediction), aimed at addressing these challenges. Methods: WSI-P2P leverages downscaled patch sampling and Multiple-Instance Learning (MIL) with transfer learning to optimize resource usage while maintaining a competitive performance. Within WSI-P2P, we introduce the K-TOP MIL aggregator, a variant of the MIL attention-based aggregator, which selectively processes the most informative K instances. The framework features an online, adaptive feature extractor that fine-tunes pre-trained models in an end-to-end manner, addressing multi-centered dataset variability. Results: WSI-P2P achieves state-of-the-art accuracy, demonstrating superior domain adaptability and computational efficiency. WSI-P2P is validated by employing several dataset splits and down-sampled patch variations, illustrating its potential as a scalable and reliable tool in clinical settings and large-scale histological studies. The framework achieved a maximum score of 95.89% AUROC and a test accuracy of 77.67% without attention, further improving to approximately 100% AUROC and a test accuracy of 95.72% recorded with the K-TOP MIL aggregator. Further, for intra-domain generalization experiments, WSI-P2P recorded a consistent performance across domains, validating its domain generalization capabilities. The K-TOP MIL aggregator also demonstrated 2.3× computational efficiency as compared to base aggregators. Conclusions: The proposed framework outperforms traditional offline feature extraction methods, ensuring high discriminative ability even when exposed to data from diverse distributions. WSI-P2P demonstrates excellent performance between subtype classifications, positioning it as a reliable tool for large-scale histological studies.

Ovarian cancer remains a significant global health concern, and its diagnosis heavily relies on whole-slide images (WSIs). Due to their gigapixel spatial resolution, WSIs must be split into patches and are usually modeled via multi-instance learning (MIL). Although previous studies have achieved remarkable performance comparable to that of humans, in clinical practice WSIs are distributed across multiple hospitals with strict privacy restrictions, necessitating secure, efficient, and effective federated MIL. Moreover, heterogeneous data distributions across hospitals lead to model heterogeneity, requiring a framework flexible to both data and model variations. This paper introduces HFed-MIL, a heterogeneous federated MIL framework that leverages gradient-based attention distillation to tackle these challenges. Specifically, we extend the intuition of Grad-CAM to the patch level and propose Patch-CAM,which computes gradient-based attention scores for each patch embedding, enabling structural knowledge distillation without explicit attention modules while minimizing privacy leakage. Beyond conventional logit distillation, we designed a dual-level objective that enforces both class-level and structural-level consistency, preventing the vanishing effect of naive averaging and enhancing the discriminative power and interpretability of the global model. Importantly, Patch-CAM scores provide a balanced solution between privacy, efficiency, and heterogeneity: they contain sufficient information for effective distillation (with minimal membership inference risk, MIA AUC ≈ 0.6) while significantly reducing communication cost (0.32 MB per round), making HFed-MIL practical for real-world federated pathology. Extensive experiments on multiple cancer subtypes and cross-domain datasets (Camelyon16, BreakHis) demonstrate that HFed-MIL achieves state-of-the-art performance with enhanced robustness under heterogeneity conditions. Moreover, the global attention visualizations yield sharper and clinically meaningful heatmaps, offering pathologists transparent insights into model decisions. By jointly balancing privacy, efficiency, and interpretability, HFed-MIL improves the practicality and trustworthiness of deep learning for ovarian cancer WSI analysis, thereby increasing its clinical significance.

Rice is a widely cultivated crop in Asia and is paramount in ensuring national and global food security. However, rice plants are susceptible to various diseases that negatively impact crop quality and quantity to meet the needs of the world’s growing population. Automated rice plant disease classification ensures food security and agricultural sustainability. Although traditional deep learning approaches have shown promising results in rice plant disease classification, the challenges posed by the heterogeneity of the data set and the feature imbalance persist. This research introduces a robust and novel Ensemble Attention Mechanism (EAM) that uses fine-tuning transfer learning to address these challenges and pre-trained (VGG16, VGG19, and customized ResNet called RiceNet, which comprises ResNet18 and ResNet50) as baseline models, specifically tailored to improve rice plant disease classification within heterogeneous datasets. The main contribution of this paper is to introduce a RiceNet framework that incorporates ensemble learning principles and attention mechanisms to adaptively balance data heterogeneity and feature representation by effectively integrating every representation to mitigate inherent class distribution imbalances. Comprehensive ablation studies validate the effectiveness of each component in the framework, demonstrating significant improvements in classification performance compared to traditional methods. Furthermore, the evaluation of RiceNet on two extensive publicly available datasets (close environment and field environment) shows its superior performance, achieving an impressive F1 score of 100

Background: Multimodal Deep learning has emerged as a crucial method for automated patient-specific quality assurance (PSQA) in radiotherapy research. Integrating image-based dose matrices with tabular plan complexity metrics enables more accurate prediction of quality indicators, including the Gamma Passing Rate (GPR) and dose difference (DD). However, modality imbalance remains a significant challenge, as tabular encoders often dominate training, suppressing image encoders and reducing model robustness. This issue becomes more pronounced under task heterogeneity, with GPR prediction relying more on tabular data, whereas dose difference prediction (DDP) depends heavily on image features. Methods: We propose BMMQA (Balanced Multi-modal Quality Assurance), a novel framework that achieves modality balance by adjusting modality-specific loss factors to control convergence dynamics. The framework introduces four key innovations: (1) task-specific fusion strategies (softmax-weighted attention for GPR regression and spatial cascading for DD prediction); (2) a balancing mechanism supported by Shapley values to quantify modality contributions; (3) a fast network forward mechanism for efficient computation of different modality combinations; and (4) a modality-contribution-based task weighting scheme for multi-task multimodal learning. A large-scale multimodal dataset comprising 1370 IMRT plans was curated in collaboration with Peking Union Medical College Hospital (PUMCH). Results: Experimental results demonstrate that, under the standard 2%/3 mm GPR criterion, BMMQA outperforms existing fusion baselines. Under the stricter 2%/2 mm criterion, it achieves a 15.7% reduction in mean absolute error (MAE). The framework also enhances robustness in critical failure cases (GPR < 90%) and achieves a peak SSIM of 0.964 in dose distribution prediction. Conclusions: Explicit modality balancing improves predictive accuracy and strengthens clinical trustworthiness by mitigating overreliance on a single modality. This work highlights the importance of addressing modality imbalance for building trustworthy and robust AI systems in PSQA and establishes a pioneering framework for multi-task multimodal learning.

Plant diseases pose significant threats to agricultural productivity and food security. Owing to a scarcity of field environment datasets, the prevailing plant disease classification approaches, trained on laboratory-controlled datasets, often grapple with achieving optimal performance in real-world environments. We proposed a novel and robust framework for Unsupervised Domain Adaptation (UDA), employing an adversarial learning approach with a Wasserstein distance-informed algorithm to learn domain invariant feature representations capable of generalizing more diverse features. This approach incorporates insights from a labeled source domain and adopts an unlabeled target domain by minimizing the distribution discrepancies between domains. Recently, mobile vision transformer (MViT)-based methods have been applied to UDA due to their ability to capture long-distance feature dependencies. However, these methods overlook the fact that MViT lacks effectiveness in extracting local feature details. The proposed framework combines the advantages of convolutional neural networks (CNNs) and MViTs, integrating local features extracted by CNNs with global features captured by MViTs. This fusion of local and global representations enhances transferability and feature discriminability within the domains. Furthermore, we incorporate a feature-fusing method to align channel dimensions and enhance the local details of the global representation. Extensive experiments using three plant disease datasets demonstrate the effectiveness and efficiency of our approach, yielding significant improvements in classification performance with 13.67%, compared to state-of-the-art (SOTA) and baseline methods. Our framework offers a promising solution for robust and efficient plant disease classification, providing valuable insights for sustainable agriculture and crop management.

Deep Learning Automated Patient-Specific Quality Assurance (PSQA) aims to reduce clinical resource requirements. It is vital to ensure the safety and effectiveness of radiation therapy by predicting the dose difference metric (Gamma passing rate) and its distribution. However, current research overlooks uncertainty quantification in model predictions, limiting their trustworthiness in real clinical environments. This paper proposes a Multi-granularity Uncertainty Quantification (MGUQ) framework. A Bayesian framework that quantifies uncertainties at multiple granularities for multi-task PSQA, specifically Gamma Passing Rate (GPR) prediction and Dose Difference Prediction (DDP), integrates visualization-based interactive components. Using Bayesian theory, we derive a comprehensive multi-granularity loss function that comprises granularity-specific loss and coherence loss components. Additionally, we proposed Multi-granularity Prior Networks, a dual-stream network architecture, to infer the distributions of DDP (modeled as t-distributions) and GPR (modeled as Gaussian distributions) under specific statistical assumptions. Comprehensive evaluations are conducted on a dataset from ''Peeking Union Medical College Hospital'', and results show that our proposed method achieves a minimum MAE loss of 0.864 with a 2%/3 mm criterion and realizes the uncertainty visualization of dose difference. Further, it also achieves 100% Clinical Accuracy (CA) with a workload of 67.2%. Experiments demonstrate that the proposed framework can enhance the trustworthiness of deep learning applications in PSQA.

Prompt Engineering has emerged as a pivotal technique in Natural Language Processing, providing a flexible approach for leveraging pre-trained language models. Particularly, a prompt is used to instruct the model to adopt the nature of given prompts, which became a well-adoptable approach in wide areas of domains. Yet, existing prompt-guided frameworks are experiencing various challenges, such as crafting prompts for specific tasks to achieve clarity and conciseness and avoid ambiguity, which requires time and computational resources. Further existing methods heavily rely on the extensive labelled datasets, yet many domain-specific challenges exist, particularly in healthcare. This study presents Prompt-Eng, a novel framework emphasizing its wide-ranging applications in healthcare, where we design precise prompts with positive and negative aspects; we hypothesize that designing prompts in pairs helps models to generalize effectively. We delve into the significance of quick design and optimization, highlighting its influence in shaping model responses. In addition, we explore the increasing demand for prompts that are aware of the context in multimodal data analysis and the incorporation of prompt engineering in new machine-learning approaches. The essence of our approach is in creating tailored prompts, which serve as instructive guidelines for the models during the prediction procedure. The proposed methodology emphasizes utilizing context-aware prompt pairs to facilitate interpreting and extracting healthcare information from a health corpus by models. The study uses the medical MIMIC-III \footnotehttps://physionet.org/content/mimiciii/1.4/ corpus to predict medicine prescriptions. The paper also explores visual and textual prompts for X-ray image analysis for pneumonia prediction on the MIMIC-CXR \footnote\urlhttps://physionet.org/content/mimic-cxr/2.0.0/ dataset. This approach stands out from existing methods by addressing challenges such as clarity, conciseness, and context awareness, thereby enabling improved interpretation and extraction of healthcare information from diverse data sources.

In recent years, large language models (LLMs) have been employed significantly in different domains of computing education. Nevertheless, these models have been focused on essential adherence to their integration as coding assistants in computing education. However, attention has been switched to thoroughly examining and analyzing LLM behavior, particularly in computing education for programming tasks such as code generation, code explanation, and programming error message explanation. Therefore, it becomes imperative to understand their behavior to examine potential pitfalls. This article addresses this gap systematically and details how different LLM-based coding chatbots, such as ChatGPT, Codex, Copilot, and others, react to various coding inputs within computing education. To achieve this objective, we collected and analyzed articles from 2021 to 2024, and 72 studies were thoroughly examined. These objectives include investigating the existing limitations and challenges associated with utilizing these systems for coding tasks, assessing their responses to prompts containing coding syntax, examining the impact of their output on student learning, and evaluating their performance as debugging tools. The findings of this review highlight that it is premature to incorporate these systems into computing education due to their limitations that may limit their effectiveness as comprehensive coding assistants for computer science students. These limitations include issues with handling prompts containing code snippets, potential negative impacts on student learning, limited debugging capabilities, and other ineffectiveness. The finding also reports multiple research directions that can be considered in future research related to LLMs in computing education.