ABSTRACT Artificial Intelligence (AI) is increasingly utilised in medicine; however, its “black‐box” nature continues to hinder clinical trust, adoption, and validation. Explainable AI (XAI) has emerged as a critical field to address these transparency challenges by making AI‐driven decisions more interpretable and actionable. This narrative review examines the progress of XAI in medicine over the past decade. We first introduce fundamental XAI concepts and describe our review methodology, followed by a comprehensive analysis of key application domains, including medical imaging, electronic health records (EHRs), and multi‐omics data. Methodologically, we categorise XAI techniques into model‐agnostic approaches (e.g., SHAP, LIME, Anchors) and model‐specific approaches (e.g., Grad‐CAM, LRP, TreeSHAP). Beyond summarising their principles, advantages, and limitations, we further provide a systematic analysis of clinical reliability and failure modes associated with each class of methods, highlighting how explanation techniques may produce misleading, unstable, or non‐causal interpretations in real‐world clinical settings. The review then discusses the demonstrated benefits of XAI, including result validation, bias detection, and improved patient–clinician communication, while critically examining persistent challenges such as limited clinical deployment, inconsistent evaluation standards, and the lack of prospective validation. Finally, we outline future research directions, emphasising the need to adapt XAI to large‐scale foundation models and conversational AI systems, as well as to extend its applicability in biomedical and multi‐omics interpretation. We argue that while XAI is essential for improving transparency, trust, and clinical adoption, its reliable and scalable integration into clinical workflows requires both methodological innovation and rigorous, clinically grounded validation frameworks.

ABSTRACT Accurate medical image segmentation is essential for disease diagnosis, treatment planning and outcome monitoring. However, current segmentation methods heavily rely on large‐scale, pixel‐level annotations, which are costly and labour‐intensive to obtain. To address this challenge, we propose TriS‐Net (Triple‐Supervision Segmentation Network) , a progressive framework that integrates image‐level, bounding box‐level and pixel‐level labels into a multigranularity supervision pipeline under limited annotation settings. In the first stage, TriS‐Net uses image‐level labels to train a classification branch, enabling the network to learn discriminative features and localise potential lesion regions. In the second stage, a box‐guided mask refinement strategy (BMR) is proposed, which combines Soft‐NMS filtering and a one‐to‐one matching mechanism to obtain reliable candidate regions. CIoU is further employed to derive image‐level quality metrics that impose quality‐aware weighted constraints on segmentation learning, thereby improving spatial localisation and structural consistency. In the third stage, a small number of pixel‐level labels are used for fine‐grained supervision, further enhancing segmentation accuracy and boundary details. The proposed method is validated on the BraTS 2019 and LiTS 2017 datasets, on which it outperforms several existing methods under limited annotation settings. Additional experiments on the BUSI ultrasound dataset further demonstrate its good generalisation capability across different imaging modalities.

Electroencephalography (EEG) emotion recognition plays a key role in improving human-machine interactions. Advanced algorithms have been proposed for this task. However, two challenges remain, i.e., unclear decision boundary in the embedded space and noise in physiological signals from various devices. To this end, we develop a novel framework, namely, UACL-Net, for EEG emotion recognition. It is based on uncertainty-aware contrastive learning (UACL) and frequency-aware self-attention (FASA). Specifically, UACL uses a multivariate Gaussian distribution to construct the latent space for different emotions. It is able to highlight interclass differences, thereby improving the robustness of model decisions. In addition, FASA generates learnable weights by applying self-attention (SA) to the real and imaginary components in the frequency domain. This helps adaptively reduce noise and capture global dependencies in temporal sequences. Our model is trained and tested on four benchmark datasets, achieving up to 94.88%, 98.71%, 96.91%, and 99.29% accuracy on SEED, DEAP, DREAMER, and FACED, respectively. Experimental results demonstrate that it is effective and has advantages over peer state-of-the-art (SOTA) methods.

As a ubiquitous portable device, mobile phones play an important role in large-scale data collection and remote health detection. Parkinson's disease (PD), a typical movement disorder, can be detected by capturing digital biomarkers using mobile phone sensors. Nevertheless, it is difficult to obtain reliable label information in large-scale remote data collection, especially for time-series digital biomarkers. Based on this, we develop a novel multi-dimensional self-supervised contrastive learning framework for remote detection of early PD by mobile phone time-series digital biomarkers. Specifically, depending on two different augmentation views, the proposed framework considers temporal contrasting, spatial contrasting, contextual contrasting, and inter-modal contrasting to enable the model to learn more discriminative features. For temporal and spatial contrasting, certain time steps (channels) of one view are used to predict the next time steps (channels) of the other view, thereby constructing a cross-view prediction task. Meanwhile, contextual contrasting is introduced into the contrast framework to consider temporal prediction and spatial prediction context representation, respectively, further increasing similarity between positive pairs and decreasing it between negative pairs. In addition, inter-modal contrasting is used to force the model to capture the potential relationship between different modal data. Experimental results show that fine-tuning with only 10% of the labeled data can outperform supervised learning and generally surpass state-of the-art algorithms.

Powered by extensive curated training data, the Segment Anything Model (SAM) demonstrates impressive generalization capabilities in open-world scenarios, effectively guided by user-provided prompts. However, the classagnostic characteristic of SAM renders its segmentation accuracy highly dependent on prompt quality. In this paper, we propose a novel plug-and-play dual-space Point Prompt Optimizer (PPO) designed to enhance prompt distribution through deep reinforcement learning (DRL)-based heterogeneous graph optimization. PPO optimizes initial prompts for any task without requiring additional training, thereby improving SAM’s downstream segmentation performance. Specifically, PPO constructs a dual-space heterogeneous graph, leveraging the robust feature-matching capabilities of a foundational pre-trained model to create internal feature and physical distance matrices. A DRL policy network iteratively refines the distribution of prompt points, optimizing segmentation predictions. We conducted experiments on four public datasets. The ablation study explores the necessity and balance of optimizing prompts in both feature and physical spaces. The comparative study shows that PPO enables SAM to surpass recent one-shot methods. Additionally, experiments with different initial prompts demonstrate PPO’s generality across prompts generated by various methods. In conclusion, PPO redefines the prompt optimization problem as a heterogeneous graph optimization task, using DRL to construct an effective, plugand-play prompt optimizer. This approach holds potential for broader applications across diverse segmentation tasks and provides a promising solution for point prompt optimization. The source code and demo are available at https://github.com/XueyuLiu/PPO.

Diagnosing Parkinson’s disease (PD) remains challenging due to its complex motor symptoms and the reliance on subjective clinical evaluations. To address this issue, this study proposes the multiscale frequency-sharing attention network (MSF-CANet), an end-to-end framework designed to identify PD and healthy control subjects using smartwatch-based inertial sensor data. MSF-CANet integrates a multiscale perception module to capture temporal features of tremors at different frequencies, a frequency-aware module to enhance PD-specific tremor signals within the 3–7 Hz range, and a shared channel attention mechanism to focus on key sensor channels while ensuring computational efficiency. The model was trained and evaluated on the PADS dataset using nested 5-fold cross-validation. The proposed method achieved an accuracy of 92.39% and an AUC of 0.9797, outperforming existing methods. The findings indicate that dual-hand data significantly improves detection performance compared to single-hand data, and dynamic tasks like “Drink from Glass” and “cross and extend both arms” achieved higher accuracy than static activities. These findings underscore the potential of MSF-CANet as a robust, noninvasive tool for real-time PD monitoring through wearable devices.

Accurately diagnosing chronic kidney disease requires pathologists to assess the structure of multiple tissues under different stains, a process that is time-consuming and labor-intensive. Current AI-based methods for automatic structure assessment, like segmentation, often demand extensive manual annotation and focus on single stain domain. To address these challenges, we introduce MSMTSeg, a generative self-supervised meta-learning framework for multi-stained multi-tissue segmentation in renal biopsy whole slide images (WSIs). MSMTSeg incorporates multiple stain transform models for style translation of inter-stain domains, a self-supervision module for obtaining pre-trained models with the domain-specific feature representation, and a meta-learning strategy that leverages generated virtual data and pre-trained models to learn the domain-invariant feature representation across multiple stains, thereby enhancing segmentation performance. Experimental results demonstrate that MSMTSeg achieves superior and robust performance, with mDSC of 0.836 and mIoU of 0.718 for multiple tissues under different stains, using only one annotated training sample for each stain. Our ablation study confirms the effectiveness of each component, positioning MSMTSeg ahead of classic advanced segmentation networks, recent few-shot segmentation methods, and unsupervised domain adaptation methods. In conclusion, our proposed few-shot cross-domain technology offers a feasible and cost-effective solution for multi-stained renal histology segmentation, providing convenient assistance to pathologists in clinical practice.

Multi-instance learning (MIL) exhibits advanced and surpassed capabilities in understanding and recognizing complex patterns within gigapixel histopathological images. However, the currents MIL methods for the analysis of the histopathological image still give rise to two main concerns. On one hand, vanilla MIL methods intuitively focus on identifying key instances (easy-to-classify) without considering hard-to-classify instances, which is biased and prone to produce false positive instance. On the other hand, since the positive tissue occupies only a small fraction of histopathological images, it is commonly suffer from class imbalance between positive and negative instances, causing the MIL model to overly focus on the majority class. In light of these issues of bias learning, we propose a multi-instance curriculum learning method that collaboratively incorporates hard negative instance mining and positive instance augmentation to improve model’s classification performance. Specifically, we first initialize the MIL model using easy-to-classify instances, then we mine the hard negative instances (hard-to-classify) and augment the positive instances via the diffusion model. Finally, the MIL model is retrained with memory rehearsal method by combining the mined negative instances and augmented positive instances. Technically, the diffusion model is first designed to generate lesion instances, which optimally augment diverse features to reflect the realistic positive samples with post screening scenario. Extensive experimental results show that the proposed method alleviates model bias in MIL and yields improvements over the state-of-the-art methods on both public datasets and private dataset.

Shaft-rate electromagnetic (EM) field is a critical feature in the detection of ships and underwater vehicles. However, the signal-to-noise ratio of the shaft-rate EM field is greatly reduced due to the presence of the static EM field, whose main energy is concentrated in the low-frequency section. In order to realize the effective detection of the weak shaft-rate EM field signals under a low signal-to-noise ratio, we propose a signal extraction method based on the Adaptive Fourier Decomposition (AFD) algorithm. At the decomposition stage, we utilize the Nevanlinna factorization and the maximal selection principle in each step, and then iteratively obtain various single components from low-frequency to high-frequency. At the extraction stage, the low-frequency information of the signal is effectively reconstructed by summing the first few components, leading to the retrieval of the shaft-rate EM field signal through residual operations. The experiment results on both synthesized and measured data show that the proposed algorithm converges faster with higher fidelity compared to the existing state-of-the-art method.

The stage prediction of kidney renal clear cell carcinoma (KIRC) is important for the diagnosis, personalized treatment, and prognosis of patients. Many prediction methods have been proposed, but most of them are based on unimodal gene data, and their accuracy is difficult to further improve. Therefore, we propose a novel multi-weighted dynamic cascade forest based on the bilinear feature extraction (MLW-BFECF) model for stage prediction of KIRC using multimodal gene data (RNA-seq, CNA, and methylation). The proposed model utilizes a dynamic cascade framework with shuffle layers to prevent early degradation of the model. In each cascade layer, a voting technique based on three gene selection algorithms is first employed to effectively retain gene features more relevant to KIRC and eliminate redundant information in gene features. Then, two new bilinear models based on the gated attention mechanism are proposed to better extract new intra-modal and inter-modal gene features; Finally, based on the idea of the bagging, a multi-weighted ensemble forest classifiers module is proposed to extract and fuse probabilistic features of the three-modal gene data. A series of experiments demonstrate that the MLW-BFECF model based on the three-modal KIRC dataset achieves the highest prediction performance with an accuracy of 88.9 %.

Abstract Classifying urine red blood cells (U‐RBCs) is the core operation in diagnosing urinary system diseases (USDs). In this paper, based on a novel data type named multi‐focus video, a multi‐instance inflated 3D convolutional neural network (MI3D) is proposed. In order to accurately classifying U‐RBCs, the MI3D integrates inflated inception‐V1 with multi‐instance learning models. Compared with the existent U‐RBC classification methods relying on single focus images, the MI3D using multi‐focus videos effectively avoids the misclassification caused by the significant deformation of U‐RBCs with the focus of microscope changing. In addition, the MI3D can learn the typical shapes and deformation patterns of U‐RBCs from multi‐focal videos simultaneously. Therefore, the accuracy of MI3D exceeds the mainstream video classification models. There are totally 597 multi‐focus videos that include four types of U‐RBCs collected to verify the effectiveness of MI3D. Experimental results show that the classification accuracy of MI3D is inspiring with 94.4%, which is obviously higher than that of existed U‐RBC classification method (85.6%). The accuracy of MI3D also achieves the comparable level with the results by junior microscopist (95.6%). Lastly, the MI3D has powerful real‐time performance, whose classification speed reaches 1.4 times than that of the microscopist.