Vision transformers (ViTs) have rapidly gained prominence in medical imaging tasks such as disease classification, segmentation, and detection due to their superior accuracy compared to conventional deep learning models. However, due to their size and complex interactions via the self-attention mechanism, they are not well understood. In particular, it is unclear whether the representations produced by such models are semantically meaningful. In this paper, using a projected gradient-based algorithm, we show that their representations are not semantically meaningful and they are inherently vulnerable to small changes. Images with imperceptible differences can have very different representations; on the other hand, images that should belong to different semantic classes can have nearly identical representations. Such vulnerability can lead to unreliable classification results; for example, unnoticeable changes cause the classification accuracy to be reduced by over 60\%. %. To the best of our knowledge, this is the first work to systematically demonstrate this fundamental lack of semantic meaningfulness in ViT representations for medical image classification, revealing a critical challenge for their deployment in safety-critical systems.

We present our solution for the Multi-Source COVID-19 Detection Challenge, which classifies chest CT scans from four distinct medical centers. To address multi-source variability, we employ the Spatial-Slice Feature Learning (SSFL) framework with Kernel-Density-based Slice Sampling (KDS). Our preprocessing pipeline combines lung region extraction, quality control, and adaptive slice sampling to select eight representative slices per scan. We compare EfficientNet and Swin Transformer architectures on the validation set. The EfficientNet model achieves an F1-score of 94.68%, compared to the Swin Transformer's 93.34%. The results demonstrate the effectiveness of our KDS-based pipeline on multi-source data and highlight the importance of dataset balance in multi-institutional medical imaging evaluation.

The use of deep learning (DL) in medical image analysis has significantly improved the ability to predict lung cancer. In this study, we introduce a novel deep convolutional neural network (CNN) model, named ResNet+, which is based on the established ResNet framework. This model is specifically designed to improve the prediction of lung cancer and diseases using the images. To address the challenge of missing feature information that occurs during the downsampling process in CNNs, we integrate the ResNet-D module, a variant designed to enhance feature extraction capabilities by modifying the downsampling layers, into the traditional ResNet model. Furthermore, a convolutional attention module was incorporated into the bottleneck layers to enhance model generalization by allowing the network to focus on relevant regions of the input images. We evaluated the proposed model using five public datasets, comprising lung cancer (LC2500 $n$=3183, IQ-OTH/NCCD $n$=1336, and LCC $n$=25000 images) and lung disease (ChestXray $n$=5856, and COVIDx-CT $n$=425024 images). To address class imbalance, we used data augmentation techniques to artificially increase the representation of underrepresented classes in the training dataset. The experimental results show that ResNet+ model demonstrated remarkable accuracy/F1, reaching 98.14/98.14\% on the LC25000 dataset and 99.25/99.13\% on the IQ-OTH/NCCD dataset. Furthermore, the ResNet+ model saved computational cost compared to the original ResNet series in predicting lung cancer images. The proposed model outperformed the baseline models on publicly available datasets, achieving better performance metrics. Our codes are publicly available at https://github.com/AIPMLab/Graduation-2024/tree/main/Peng.

This study aimed to develop a pretreatment CT-based multichannel predictor integrating deep learning features encoded by Transformer models for preoperative diagnosis of major pathological response (MPR) in non-small cell lung cancer (NSCLC) patients receiving neoadjuvant immunochemotherapy. This multicenter diagnostic study retrospectively included 332 NSCLC patients from four centers. Pretreatment computed tomography images were preprocessed and segmented into region of interest cubes for radiomics modeling. These cubes were cropped into four groups of 2 dimensional image modules. GoogLeNet architecture was trained independently on each group within a multichannel framework, with gradient-weighted class activation mapping and SHapley Additive exPlanations value‌ for visualization. Deep learning features were carefully extracted and fused across the four image groups using the Transformer fusion model. After models training, model performance was evaluated via the area under the curve (AUC), sensitivity, specificity, F1 score, confusion matrices, calibration curves, decision curve analysis, integrated discrimination improvement, net reclassification improvement, and DeLong test. The dataset was allocated into training (n = 172, Center 1), internal validation (n = 44, Center 1), and external test (n = 116, Centers 2-4) cohorts. Four optimal deep learning models and the best Transformer fusion model were developed. In the external test cohort, traditional radiomics model exhibited an AUC of 0.736 [95% confidence interval (CI): 0.645-0.826]. The‌ optimal deep learning imaging ‌module‌ showed superior AUC of 0.855 (95% CI: 0.777-0.934). The fusion model named Transformer_GoogLeNet further improved classification accuracy (AUC = 0.924, 95% CI: 0.875-0.973). The new method of fusing multichannel deep learning with the Transformer Encoder can accurately diagnose whether NSCLC patients receiving neoadjuvant immunochemotherapy will achieve MPR. Our findings may support improved surgical planning and contribute to better treatment outcomes through more accurate preoperative assessment.

To develop and validate an ensemble machine learning ultrasound radiomics model for predicting drug resistance in lymph node tuberculosis (LNTB). This multicenter study retrospectively included 234 cervical LNTB patients from one center, randomly divided into training (70%) and internal validation (30%) cohorts. Radiomic features were extracted from ultrasound images, and an L1-based method was used for feature selection. A predictive model combining ensemble machine learning and AdaBoost algorithms was developed to predict drug resistance. Model performance was assessed using independent external test sets (Test A and Test B) from two other centres, with metrics including AUC, accuracy, precision, recall, F1 score, and decision curve analysis. Of the 851 radiometric features extracted, 161 were selected for the model. The model achieved AUCs of 0.998 (95% CI: 0.996-0.999), 0.798 (95% CI: 0.692-0.904), 0.846 (95% CI: 0.700-0.992), and 0.831 (95% CI: 0.688-0.974) in training, internal validation, and external test sets A and B, respectively. The decision curve analysis showed a substantial net benefit across a threshold probability range of 0.38 to 0.57. The LNTB resistance prediction model developed demonstrated high diagnostic efficacy in both internal and external validation. Radiomics, through the application of ensemble machine learning algorithms, provides new insights into drug resistance mechanisms and offers potential strategies for more effective patient treatment. Lymph node tuberculosis; Drug resistance; Ultrasound; Radiomics; Machine learning.

Surgical resection is an effective treatment for medically refractory mesial temporal lobe epilepsy (mTLE), however, more than one-third of patients fail to achieve seizure freedom after surgery. This study aimed to evaluate preoperative individual morphometric network characteristics and develop a machine learning model to predict surgical outcome in mTLE. This multicentre, retrospective study included 189 mTLE patients who underwent unilateral temporal lobectomy and 78 normal controls between February 2018 and June 2023. Postoperative seizure outcomes were categorized as seizure-free (SF, n = 125) or non-seizure-free (NSF, n = 64) at a minimum of one-year follow-up. The preoperative individualized structural covariance network (iSCN) derived from T1-weighted MRI was constructed for each patient by calculating deviations from the control-based reference distribution, and further divided into the surgery network and the surgically spared network using a standard resection mask by merging each patient's individual lacuna. Regional features were selected separately from bilateral, ipsilateral and contralateral iSCN abnormalities to train support vector machine models, validated in two independent external datasets. NSF patients showed greater iSCN deviations from the normative distribution in the surgically spared network compared to SF patients (P = 0.02). These deviations were widely distributed in the contralateral functional modules (P < 0.05, false discovery rate corrected). Seizure outcome was optimally predicted by the contralateral iSCN features, with an accuracy of 82% (P < 0.05, permutation test) and an area under the receiver operating characteristic curve (AUC) of 0.81, with the default mode and fronto-parietal areas contributing most. External validation in two independent cohorts showed accuracy of 80% and 88%, with AUC of 0.80 and 0.82, respectively, emphasizing the generalizability of the model. This study provides reliable personalized structural biomarkers for predicting surgical outcome in mTLE and has the potential to assist tailored surgical treatment strategies.

To address misdiagnosis caused by feature coupling in multi-label medical image classification, this study introduces a chest X-ray pathology reasoning method. It combines hierarchical attention convolutional networks with a multi-label decoupling loss function. This method aims to enhance the precise identification of complex lesions. It dynamically captures multi-scale lesion morphological features and integrates lung field partitioning with lesion localization through a dual-path attention mechanism, thereby improving clinical disease prediction accuracy. An adaptive dilated convolution module with 3 × 3 deformable kernels dynamically captures multi-scale lesion features. A channel-space dual-path attention mechanism enables precise feature selection for lung field partitioning and lesion localization. Cross-scale skip connections fuse shallow texture and deep semantic information, enhancing microlesion detection. A KL divergence-constrained contrastive loss function decouples 14 pathological feature representations via orthogonal regularization, effectively resolving multi-label coupling. Experiments on ChestX-ray14 show a weighted F1-score of 0.97, Hamming Loss of 0.086, and AUC values exceeding 0.94 for all pathologies. This study provides a reliable tool for multi-disease collaborative diagnosis.

Alzheimer's disease (AD) constitutes a neurodegenerative disorder predominantly observed in the geriatric population. If AD can be diagnosed early, both in terms of prevention and treatment, it is very beneficial to patients. Therefore, our team proposed a novel deep learning model named 3D-CNN-VSwinFormer. The model consists of two components: the first part is a 3D CNN equipped with a 3D Convolutional Block Attention Module (3D CBAM) module, and the second part involves a fine-tuned Video Swin Transformer. Our investigation extracts features from subject-level 3D Magnetic resonance imaging (MRI) data, retaining only a single 3D MRI image per participant. This method circumvents data leakage and addresses the issue of 2D slices failing to capture global spatial information. We utilized the ADNI dataset to validate our proposed model. In differentiating between AD patients and cognitively normal (CN) individuals, we achieved accuracy and AUC values of 92.92% and 0.9660, respectively. Compared to other studies on AD and CN recognition, our model yielded superior results, enhancing the efficiency of AD diagnosis.

Endobronchial ultrasound (EBUS) is a widely used imaging modality for evaluating thoracic lymph nodes (LNs), particularly in the staging of lung cancer. Artificial intelligence (AI)-assisted EBUS has emerged as a promising tool to enhance diagnostic accuracy. However, its effectiveness in differentiating benign from malignant thoracic LNs remains uncertain. This meta-analysis aimed to evaluate the diagnostic performance of AI-assisted EBUS compared to the pathological reference standards. A systematic search was conducted across PubMed, Embase, and Web of Science for studies assessing AI-assisted EBUS in differentiating benign and malignant thoracic LNs. The reference standard included pathological confirmation via EBUS-guided transbronchial needle aspiration, surgical resection, or other histological/cytological validation methods. Sensitivity, specificity, diagnostic likelihood ratios, and diagnostic odds ratio (OR) were pooled using a random-effects model. The area under the receiver operating characteristic curve (AUROC) was summarized to evaluate diagnostic accuracy. Subgroup analyses were conducted by study design, lymph node location, and AI model type. Twelve studies with a total of 6,090 thoracic LNs were included. AI-assisted EBUS showed a pooled sensitivity of 0.75 (95% confidence interval [CI]: 0.60-0.86, I² = 97%) and specificity of 0.88 (95% CI: 0.83-0.92, I² = 96%). The positive and negative likelihood ratios were 6.34 (95% CI: 4.41-9.08) and 0.28 (95% CI: 0.17-0.47), respectively. The pooled diagnostic OR was 22.38 (95% CI: 11.03-45.38), and the AUROC was 0.90 (95% CI: 0.88-0.93). The subgroup analysis showed higher sensitivity but lower specificity in retrospective studies compared to prospective ones (sensitivity: 0.87 vs. 0.42; specificity: 0.80 vs. 0.93; both p < 0.001). No significant differences were found by lymph node location or AI model type. AI-assisted EBUS shows promise in differentiating benign from malignant thoracic LNs, particularly those with high specificity. However, substantial heterogeneity and moderate sensitivity highlight the need for cautious interpretation and further validation. PROSPERO CRD42025637964.

The lack of reliable biomarkers for the early detection and risk stratification of post-COVID-19 pulmonary fibrosis (PCPF) underscores the urgency advanced predictive tools. This study aimed to develop a machine learning-based predictive model integrating quantitative CT (qCT) radiomics and clinical features to assess the risk of lung fibrosis in COVID-19 patients. A total of 204 patients with confirmed COVID-19 pneumonia were included in the study. Of these, 93 patients were assigned to the development cohort (74 for training and 19 for internal validation), while 111 patients from three independent hospitals constituted the external validation cohort. Chest CT images were analyzed using qCT software. Clinical data and laboratory parameters were obtained from electronic health records. Least absolute shrinkage and selection operator (LASSO) regression with 5-fold cross-validation was used to select the most predictive features. Twelve machine learning algorithms were independently trained. Their performances were evaluated by receiver operating characteristic (ROC) curves, area under the curve (AUC) values, sensitivity, and specificity. Seventy-eight features were extracted and reduced to ten features for model development. These included two qCT radiomics signatures: (1) whole lung_reticulation (%) interstitial lung disease (ILD) texture analysis, (2) interstitial lung abnormality (ILA)_Num of lung zones ≥ 5%_whole lung_ILA. Among 12 machine learning algorithms evaluated, the support vector machine (SVM) model demonstrated the best predictive performance, with AUCs of 0.836 (95% CI: 0.830-0.842) in the training cohort, 0.796 (95% CI: 0.777-0.816) in the internal validation cohort, and 0.797 (95% CI: 0.691-0.873) in the external validation cohort. The integration of CT radiomics, clinical and laboratory variables using machine learning provides a robust tool for predicting pulmonary fibrosis progression in COVID-19 patients, facilitating early risk assessment and intervention.