In the cognitive process of histopathology-based cancer diagnosis, pathologists form an initial fuzzy abnormality judgment through global visual perception, and then perform detailed multi-scale evaluation, which leads to diagnosis under the guidance of potential disease associations. However, existing diagnosis methods rarely formalize this pathologist-inspired process, which limits diagnostic-process transparency and predictive reliability. Consequently, we propose a fuzzy cognitive mimicking network with reasoning agent (FCMNet-RA) for histopathology-based cancer diagnosis. The network constructs a fuzzy abnormality perception module to model the initial fuzzy abnormality judgment. The module extracts multi-granularity information from different stages and quantifies abnormal cues through fuzzy membership functions, thereby forming fuzzy perception of potential abnormal regions. To further organize multi-scale evaluation, the network constructs a stage-wise deliberative reasoning agent. The reasoning agent designs stage adapters and a cognition-sharing subnetwork for unified evaluation, and then regulates the contribution of each stage to the final diagnosis according to fuzzy abnormality cues and stage information. Moreover, a cognitive prototype calibration mechanism is designed to incorporate the guidance of potential disease associations. The mechanism constructs disease cognitive prototypes with diagnostic logic priors to encode potential category associations inspired by pathological diagnosis, thereby guiding different network stages toward consistent representations aligned with the final diagnosis. Overall, FCMNet-RA adopts a pathologist-inspired process design to improve diagnostic-process transparency and predictive reliability. Comprehensive experiments on two private clinical datasets and two public datasets demonstrate that FCMNet-RA outperforms existing diagnosis methods, while qualitative and quantitative analyses provide further support for diagnostic-process transparency and predictive reliability.
Emotion recognition based on electroencephalograms (EEGs) is crucial for affective brain-computer interfaces. However, sparse labeled data and domain distribution inconsistencies resulting from temporal non-stationarity and inter-subject variability significantly impede its practical deployment. To address these issues, we propose a two-stage framework called SCPMDA (Soft Contrastive Pre-training and Multi-source Domain Adaptation). In the first stage, a self-supervised pre-training strategy is designed to learn transferable EEG representations from unlabeled data. Specifically, a MoCo-based soft contrastive objective is adopted to mitigate false-negative interference, while mutual information regularization enhances representation consistency across augmented views. To effectively reduce cross-domain disparities, we present a multi-source-domain adaptive framework in the second stage that incorporates the pre-training strategy. The framework adaptively leverages multiple source domains while mitigating negative transfer by utilizing a dual-channel attention mechanism to capture discriminative spatial–spectral features and implementing a source-domain competition strategy based on Wasserstein distance and dynamically weighted MMD. In addition, high-confidence pseudo-label learning and class-level alignment are introduced to improve conditional distribution matching. The proposed SCPMDA model was evaluated using the SEED and SEED IV datasets, achieving mean accuracies of 93.45% and 82.97% in cross-subject experiments as well as 96.50% and 86.33% in cross-session experiments, respectively. These results indicate the potential effectiveness of SCPMDA for EEG emotion recognition under cross-subject and cross-session domain adaptation settings.
As the core of Traditional Chinese Medicine (TCM) diagnosis and treatment, herb recommendation aims to recommend a group of herbs according to patients’ symptoms. Existing graph-based methods mainly model pairwise relations and either omit explicit syndrome and state-element knowledge or represent the common induction of multiple patients through independent pairwise edges, thereby failing to model it as a single high-order relation. To address this limitation, we aim to improve herb recommendation by explicitly modeling the high-order patient relations induced by shared syndromes and state-element configurations. We propose a multi-hypergraph convolutional neural network (MHGCN) that represents each patient as a node and constructs a syndrome-induced patient hypergraph and a state-element-induced patient hypergraph. In each hypergraph, one hyperedge connects all patients sharing the same syndrome or exact state-element configuration. Parallel encoders learn from the two hypergraphs, and their representations are fused for multi-label herb prediction using a frequency-weighted binary cross-entropy loss. Across ten runs, MHGCN achieved mean F1-score@5 values of 70.532% on the Treatise on Febrile Diseases dataset and 29.737% on the Dictionary of Traditional Chinese Medicine Prescriptions dataset, respectively. It outperformed all evaluated topic-model, graph-based, and large language model baselines. In paired analyses using matched random seeds, its improvements over LightGCNM in Precision@5 and F1-score@10 remained significant under both paired t-tests and Wilcoxon signed-rank tests after Holm correction (padj<0.05). These findings demonstrate that representing each common induction event as a patient hyperedge enables MHGCN to model syndrome- and state-element-induced high-order relations directly and supports its effectiveness for herb recommendation.
Specific Language Impairment (SLI), or developmental dysphasia, affects 2%–11% of children and poses significant risks to academic success, communication, and social development. While early diagnosis is critical for effective intervention, the heterogeneity of symptoms and a lack of objective tools often lead to misdiagnosis. This work introduces an automated SLI detection system based on a novel dynamic mode Hilbert transform (DMHT) framework. The proposed method first applies dynamic mode decomposition (DMD) to speech signals to extract their constituent dynamic modes. Each mode is then analyzed using the Hilbert transform (HT) to derive its instantaneous amplitude and instantaneous frequency components. From these temporal components, a set of statistical features such as mean, variance, skewness, kurtosis, and entropy measures, including approximate, permutation, and increment entropy, is computed to form a discriminative feature set. These DMHT-based features are designed to capture the complex, non-linear dynamical signatures of SLI in speech. The proposed technique achieves a maximum accuracy of 97.20% for the vowel/i/ using a random forest (RF) classifier. The results show that the DMHT pipeline could be a reliable, feature-rich approach for automatically and objectively detecting SLI from speech signals.
In breast ultrasound segmentation, the accurate delineation of lesions is often severely hindered by pervasive background noise and uninformative feature responses. Existing methods typically struggle with feature redundancy during propagation, where irrelevant background information overshadows critical structures. This not only degrades segmentation performance but also incurs unnecessary computational overhead. To address this bottleneck, we propose SFPNet, a Sparsity-Driven Feature Purification Network designed to transition representations from redundancy to purity. Specifically, the Multi-Scale Dynamic Hybrid Modeling (MDHM) module is designed to comprehensively capture diverse anatomical structures and extensive multi-scale features. These representations are subsequently calibrated to evaluate feature importance and perform an initial suppression of redundancy. Secondly, we introduce an Adaptive Cross-Scale Purification Strategy (ACPS) that acts as the core purifier in the network. Driven by sparsity-guided modulation and cross-scale aggregation, ACPS helps suppress redundant responses and preserve informative structural features for decoder reconstruction. Extensive experiments are conducted on five public datasets, including three breast ultrasound datasets and two non-breast medical datasets. The results demonstrate that SFPNet achieves competitive segmentation performance while maintaining a favorable efficiency-performance balance. These findings suggest that SFPNet offers a promising lightweight framework for breast ultrasound lesion segmentation by balancing segmentation accuracy and computational efficiency. The source code of SFPNet will be publicly available at https://github.com/ccpipi/SFPNet.
