Automated segmentation of mediastinal neoplasms with preoperative computed tomography (CT) scans is critical for clinical diagnosis. Although convolutional neural networks (CNNs) have proven effective in medical imaging segmentation, the segmentation of mediastinal neoplasms, which vary greatly in shape, size, and texture, presents a unique challenge due to the inherent local focus of convolution operations. To address this limitation, we propose a confidence-enhanced semi-supervised learning framework for mediastinal neoplasm segmentation. Specifically, we introduce a confidence-enhanced module that improves segmentation accuracy over indistinct tumor boundaries by assessing and excluding unreliable predictions simultaneously, which can greatly enhance the efficiency of exploiting unlabeled data. In addition, we implement an iterative learning strategy designed to continuously refine prediction reliability estimates throughout the training process, ensuring more precise confidence assessments. Quantitative analysis on a real-world dataset demonstrates that our model significantly improves the performance by leveraging unlabeled data, surpassing existing semi-supervised segmentation benchmarks. Finally, to promote more efficient academic communication, the analysis code is available at https://github.com/fxiaotong432/CEDS

With the rise of online health communities (OHC), predicting patient emotional attitudes based on doctor-patient dialogue texts has emerged as a critical focus of research. Unlike traditional sentiment analysis, OHC texts exhibit two distinct characteristics: (1) they originate from both participants in a dialogue rather than a single individual, and (2) patient satisfaction depends not only on the consultation process, but also on the doctor's personal attributes. Currently, there is a lack of benchmark datasets that simultaneously address both of these aspects to predict patient satisfaction in OHC scenarios. To bridge this gap, this paper presents a large-scale, dual-modality dataset DialogMedTab based on multi-turn doctor-patient dialogues and introduces a Cross-Decoder Transformer (CDT) as the baseline model. Key innovations of CDT include: (1) an Agent Embedding to encode speaker roles (e.g., doctor vs. patient) and (2) Cluster Distance-Based Embedding (CDB Embedding) for numerical feature representation. (3) A cross-decoder architecture achieves layer-wise fusion of textual and tabular modalities, and (4) the proposed dual-alignment loss ensures robust modality alignment prior to cross-modal fusion. We evaluated the performance of CDT against four baseline models and two large language models (LLMs) on DialogMedTab and three additional public datasets. Experimental results show that CDT outperforms the baseline models by 0.1 % to 3.3% in accuracy and CDT-BERT also exceeds fine-tuned LLMs on most datasets. The dataset and the code are publicly available at https://huggingface.co/datasets/FireflyLiu/DialogMedTab and https://github.com/RUCJing/Cross_Decoder_Transformer.

Chronic respiratory diseases (CRDs) are among the leading chronic diseases worldwide and pose a significant threat to public health and life expectancy. Pulmonary dysfunction, a core manifestation of CRDs, is routinely assessed using the pulmonary function test (PFT). Although PFT is essential for the evaluation and diagnosis of CRDs, its widespread adoption and patient compliance remain limited. Recent studies have demonstrated strong correlations between chest CT imaging features and PFT metrics, suggesting a viable alternative approach for assessing pulmonary function. To address this issue, we develop a copula-augmented lightweight deep learning model (CALM) for robust pulmonary function assessment from CT imaging. Specifically, our study centers on three critical pulmonary function indices: forced vital capacity (FVC), forced expiratory volume in 1 second (FEV1), and total lung capacity (TLC). The proposed CALM framework integrates a multi-view guided attention mechanism to extract discriminative spatial features from CT images. Unlike previous research, we introduce a copula-based loss function specifically designed to model the conditional dependencies among pulmonary function index parameters. Lastly, experiments in a real-world hospital data set indicate that CALM achieves better prediction accuracy compared to baseline methods, while requiring fewer trainable parameters. In addition to providing a novel approach for the screening and diagnosis of CRDs, the proposed method also advances a lightweight, copula-enhanced joint modeling framework from a technical perspective. The code is available at https://github.com/xy015/CALM.

Automatic segmentation and classification of thymic tumors based on preoperative CT scans are critical for clinical diagnosis. However, significant variability in the shape, size, and texture of thymic tumors, along with their blurred boundaries and complex pathological features, poses substantial challenges to automated recognition. This study focuses on two key objectives: (1) semantic segmentation at the pixel level of thymic tumors on CT images and (2) identification of high-risk thymic carcinoma. To address the above challenges, we propose a multiview attention-guided network (MAG-Net), a novel multitask learning framework guided by attention mechanisms. The model simultaneously takes 3D subvolumes of equal size extracted from axial, coronal, and sagittal views of the CT scan as input and fuses features across multiple views under the guidance of attention mechanisms. Moreover, we introduce a segmentation-classification prior attention (SCPA) module that embeds spatial location cues from segmentation into feature learning for classification. Extensive experiments conducted on both our own collected data set and public data sets demonstrate the effectiveness of the proposed method, achieving a dice coefficient of 90.54% for the segmentation task and an AUC of nearly 0.9 for classification. To facilitate further research, the analysis code is available on https://github.com/weixuxuxu/MAGNet.

Lung cancer remains a leading cause of cancer-related mortality, with early detection of pulmonary nodules critical for improving patient outcomes. Although deep learning models have shown promise in nodule detection using low-dose CT scans, existing methods struggle with generalization across diverse nodule morphologies, particularly for micronodules (≤10mm). To address these challenges, we propose DSANet, a novel 3D Deformable Slice-Aware Network featuring an adaptive deformable slice grouped (DSG) module. The DSG module dynamically adjusts slice grouping strategies and attention weights based on CT image features, enhancing 3D spatial feature extraction for nodules of varying sizes and shapes. We evaluated DSANet on both our large-scale ChestCT2025 dataset (1,000 scans, 1,465 nodules) and the public LUNA2016 dataset. Experimental results demonstrate that DSANet outperforms state-of-the-art methods in key metrics, with significant improvements in small-nodule detection. Ablation studies confirm the critical role of the DSG module in increasing the detection accuracy. Our approach offers a robust solution for early lung cancer diagnosis, particularly in challenging micronodule cases. The code is available at https://github.com/czy020202/DSANet.

As a common subtype of lung cancer, the diagnosis of lung adenocarcinoma has significant importance in clinical practice, particularly in distinguishing between pre-invasive adenocarcinoma (Pre-IA) and invasive adenocarcinoma (IAC). This distinction is critical because the two types of lesions correspond to different clinical treatment strategies: Pre-IAs typically require only regular observation, while IACs necessitate immediate surgical removal. In this article, we propose a novel deep learning model, the Radiomic Feature Deep Factorization Machine (RFDFM) network model, for distinguishing IACs from Pre-IAs in CT images, leveraging both radiomic features and deep learning features. Our novelty resides in pioneering the application of recommendation system model structures to computer-aided diagnosis of pulmonary nodules, demonstrating feasibility and effectively addressing the limitations of traditional methods in handling radiomic features. Moreover, the use of low-level feature fusion convolutional neural networks minimizes the information loss, and an element-wise attention mechanism in feature fusion stage to accentuate key features and improve model fitting. For extensive validation, 1,052 nodule samples were gathered from a total of 791 patients that were diagnosed with lung adenocarcinoma across two top-tier hospitals. The proposed RFDFM method can achieve a sota performance of 94.2% in terms of AUC. Results of extensive ablation studies demonstrate its contribution to improved performance. Finally, to promote more efficient academic communication, the analysis code is publicly available at https://github.com/Chengcheng-Guo/RFDFM.

We study here a Gaussian mixture model (GMM) with rare events data. In this case, the commonly used Expectation-Maximization (EM) algorithm exhibits extremely slow numerical convergence rate. To theoretically understand this phenomenon, we formulate the numerical convergence problem of the EM algorithm with rare events data as a problem about a contraction operator. Theoretical analysis reveals that the spectral radius of the contraction operator in this case could be arbitrarily close to 1 asymptotically. This theoretical finding explains the empirical slow numerical convergence of the EM algorithm with rare events data. To overcome this challenge, a Mixed EM (MEM) algorithm is developed, which utilizes the information provided by partially labeled data. As compared with the standard EM algorithm, the key feature of the MEM algorithm is that it requires additionally labeled data. We find that MEM algorithm significantly improves the numerical convergence rate as compared with the standard EM algorithm. The finite sample performance of the proposed method is illustrated by both simulation studies and a real-world dataset of Swedish traffic signs.

Lung cancer, particularly in its advanced stages, remains a leading cause of death globally. Though early detection via low-dose computed tomography (CT) is promising, the identification of high-risk factors crucial for surgical mode selection remains a challenge. Addressing this, our study introduces an Attention-Enhanced Graph Convolutional Network (AE-GCN) model to classify whether there are high-risk factors in stage I lung cancer based on the preoperative CT images. This will aid surgeons in determining the optimal surgical method before the operation. Unlike previous studies that relied on 3D patch techniques to represent nodule spatial features, our method employs a GCN model to capture the spatial characteristics of pulmonary nodules. Specifically, we regard each slice of the nodule as a graph vertex, and the inherent spatial relationships between slices form the edges. Then, to enhance the expression of nodule features, we integrated both channel and spatial attention mechanisms with a pre-trained VGG model for adaptive feature extraction from pulmonary nodules. Lastly, the effectiveness of the proposed method is demonstrated using real-world data collected from the hospitals, thereby emphasizing its potential utility in the clinical practice.

This paper addresses the emerging task of recognizing multiple retinal diseases from wide-field (WF) and ultra-wide-field (UWF) fundus images. For an effective use of existing large amount of labeled color fundus photo (CFP) data and the relatively small amount of WF and UWF data, we propose a supervised domain adaptation method named Cross-domain Collaborative Learning (CdCL). Inspired by the success of fixed-ratio based mixup in unsupervised domain adaptation, we re-purpose this strategy for the current task. Due to the intrinsic disparity between the field-of-view of CFP and WF/UWF images, a scale bias naturally exists in a mixup sample that the anatomic structure from a CFP image will be considerably larger than its WF/UWF counterpart. The CdCL method resolves the issue by Scale-bias Correction, which employs Transformers for producing scale-invariant features. As demonstrated by extensive experiments on multiple datasets covering both WF and UWF images, the proposed method compares favorably against a number of competitive baselines.

The judgment of controversial cases has always been an important judicial issue, but it is not easy to discover them in practice.In this paper, based on 1,361,354 legal instruments data collected from China Judgments Online, we adopt a deep learning framework to classify 147 different kinds of crimes.The proposed method has three critical steps: 1) We adopt a deep learning model to predict crime categorization; 2) With the trained model, each case is given a score vector which represents the probability that it belongs to each crime; 3) With the probability score, we develop an entropy-based index to measure the controversy of each case.We find that the larger the entropy, the more inconsistent the result given by the model based on the first instance judgment.To verify the proposed entropy measure, we provide 1) two-sided evidence based on second instance judgments; 2) comparison with some baseline models.Both confirm the practical usefulness of the entropy measure.Our results indicate that the proposed framework has an ability to discover potentially controversial cases.It should be noted that the goal of this study is not to substitute the model result for the judge's decision, but to provide a guiding reference for the judicial practice of sentencing.