Despite significant progress in applying large language models (LLMs) to the medical domain, several limitations still prevent them from practical applications. Among these are the constraints on model size and the lack of cohort-specific labeled datasets. In this work, we investigated the potential of improving a lightweight LLM, such as Llama 3.1-8B, through fine-tuning with datasets using synthetic labels. Two tasks are jointly trained by combining their respective instruction datasets. When the quality of the task-specific synthetic labels is relatively high (e.g., generated by GPT4-o), Llama 3.1-8B achieves satisfactory performance on the open-ended disease detection task, with a micro F1 score of 0.91. Conversely, when the quality of the task-relevant synthetic labels is relatively low (e.g., from the MIMIC-CXR dataset), fine-tuned Llama 3.1-8B is able to surpass its noisy teacher labels (micro F1 score of 0.67 v.s. 0.63) when calibrated against curated labels, indicating the strong inherent underlying capability of the model. These findings demonstrate the potential offine-tuning LLMs with synthetic labels, offering a promising direction for future research on LLM specialization in the medical domain.

This work aims to promote early and accurate diagnosis of Temporal Lobe Epilepsy (TLE) by developing state-of-the-art deep learning techniques, with the goal of minimizing the consequences of epilepsy on individuals and society. Current approaches for TLE detection have drawbacks, including applicability to particular MRI sequences, moderate ability to determine the side of the onset zones, and weak cross-validation with different patient groups, which hampers their practical use. To overcome these difficulties, a new Hybrid Attention-Enhanced Transformer Network (HAETN) is introduced for early TLE diagnosis. This approach uses newly developed Fuzzy-AAL Segmentation Framework (FASF) which is a combination of Fuzzy Possibilistic C-Means (FPCM) algorithm for segmentation of tissue and AAL labelling for labelling of tissues. Furthermore, an effective feature selection method is proposed using the Dipper- grey wolf optimization (DGWO) algorithm to improve the performance of the proposed model. The performance of the proposed method is thoroughly assessed by accuracy, sensitivity, and F1-score. The performance of the suggested approach is evaluated on the Temporal Lobe Epilepsy-UNAM MRI Dataset, where it attains an accuracy of 98.61%, a sensitivity of 99.83%, and F1-score of 99.82%, indicating its efficiency and applicability in clinical practice.

To enhance glioma segmentation, a 3D-MRI intelligent glioma segmentation method based on deep learning is introduced. This method offers significant guidance for medical diagnosis, grading, and treatment strategy selection. Glioma case data were sourced from the BraTS2023 public dataset. Firstly, we preprocess the dataset, including 3D clipping, resampling, artifact elimination and normalization. Secondly, in order to enhance the perception ability of the network to different scale features, we introduce the space pyramid pool module. Then, by making the model focus on glioma details and suppressing irrelevant background information, we propose a multi-scale fusion attention mechanism; And finally, to address class imbalance and enhance learning of misclassified voxels, a combination of Dice and Focal loss functions was employed, creating a loss function, this method not only maintains the accuracy of segmentation, It also improves the recognition of challenge samples, thus improving the accuracy and generalization of the model in glioma segmentation. Experimental findings reveal that the enhanced 3D U-Net network model stabilizes training loss at 0.1 after 150 training iterations. The refined model demonstrates superior performance with the highest DSC, Recall, and Precision values of 0.7512, 0.7064, and 0.77451, respectively. In Whole Tumor (WT) segmentation, the Dice Similarity Coefficient (DSC), Recall, and Precision scores are 0.9168, 0.9426, and 0.9375, respectively. For Core Tumor (TC) segmentation, these scores are 0.8954, 0.9014, and 0.9369, respectively. In Enhanced Tumor (ET) segmentation, the method achieves DSC, Recall, and Precision values of 0.8674, 0.9045, and 0.9011, respectively. The DSC, Recall, and Precision indices in the WT, TC, and ET segments using this method are the highest recorded, significantly enhancing glioma segmentation. This improvement bolsters the accuracy and reliability of diagnoses, ultimately providing a scientific foundation for clinical diagnosis and treatment.

Magnetic Resonance Imaging (MRI) based three-dimensional analysis of knee cartilage has evolved to become fully automatic. However, when implementing these measurements across multiple clinical centers, scanner variability becomes a critical consideration. Our purposes were to quantify and compare same-model variability (between repeated scans on the same MRI system) and cross-model variability (across different MRI systems) in knee cartilage thickness measurements using MRI scanners from five manufacturers, as analyzed with a specific 3D volume analysis software. Ten healthy volunteers (eight males and two females, aged 22-60 years) underwent two scans of their right knee on 3T MRI systems from five manufacturers (Canon, Fujifilm, GE, Philips, and Siemens). The imaging protocol included fat-suppressed spoiled gradient echo and proton density weighted sequences. Cartilage regions were automatically segmented into 7 subregions using a specific deep learning-based 3D volume analysis software. This resulted in 350 measurements for same-model variability and 2,800 measurements for cross-model variability. For same-model variability, 82% of measurements showed variability ≤0.10 mm, and 98% showed variability ≤0.20 mm. For cross-model variability, 51% showed variability ≤0.10 mm, and 84% showed variability ≤0.20 mm. The mean same-model variability (0.06 ± 0.05 mm) was significantly lower than cross-model variability (0.11 ± 0.09 mm) (p < 0.001). This study demonstrates that knee cartilage thickness measurements exhibit significantly higher variability across different MRI systems compared to repeated measurements on the same system, when analyzed using this specific software. This finding has important implications for multi-center studies and longitudinal assessments using different MRI systems and highlights the software-dependent nature of such variability assessments.

Early and accurate detection of COVID-19 and pneumonia through medical imaging is critical for effective patient management. This study aims to develop a robust framework that integrates synthetic image augmentation with advanced deep learning (DL) models to address dataset imbalance, improve diagnostic accuracy, and enhance trust in artificial intelligence (AI)-driven diagnoses through Explainable AI (XAI) techniques. The proposed framework benchmarks state-of-the-art models (InceptionV3, DenseNet, ResNet) for initial performance evaluation. Synthetic images are generated using Feature Interpolation through Linear Mapping and principal component analysis to enrich dataset diversity and balance class distribution. YOLOv8 and InceptionV3 models, fine-tuned via transfer learning, are trained on the augmented dataset. Grad-CAM is used for model explainability, while large language models (LLMs) support visualization analysis to enhance interpretability. YOLOv8 achieved superior performance with 97% accuracy, precision, recall, and F1-score, outperforming benchmark models. Synthetic data generation effectively reduced class imbalance and improved recall for underrepresented classes. Comparative analysis demonstrated significant advancements over existing methodologies. XAI visualizations (Grad-CAM heatmaps) highlighted anatomically plausible focus areas aligned with clinical markers of COVID-19 and pneumonia, thereby validating the model's decision-making process. The integration of synthetic data generation, advanced DL, and XAI significantly enhances the detection of COVID-19 and pneumonia while fostering trust in AI systems. YOLOv8's high accuracy, coupled with interpretable Grad-CAM visualizations and LLM-driven analysis, promotes transparency crucial for clinical adoption. Future research will focus on developing a clinically viable, human-in-the-loop diagnostic workflow, further optimizing performance through the integration of transformer-based language models to improve interpretability and decision-making.

Thyroid nodule, as a common clinical endocrine disease, has become increasingly prevalent worldwide. Ultrasound, as the premier method of thyroid imaging, plays an important role in accurately diagnosing and managing thyroid nodules. However, there is a high degree of inter- and intra-observer variability in image interpretation due to the different knowledge and experience of sonographers who have huge ultrasound examination tasks everyday. Artificial intelligence based on computer-aided diagnosis technology maybe improve the accuracy and time efficiency of thyroid nodules diagnosis. This study introduced an artificial intelligence software called SW-TH01/II to evaluate ultrasound image characteristics of thyroid nodules including echogenicity, shape, border, margin, and calcification. We included 225 ultrasound images from two hospitals in Shanghai, respectively. The sonographers and software performed characteristics analysis on the same group of images. We analyzed the consistency of the two results and used the sonographers' results as the gold standard to evaluate the accuracy of SW-TH01/II. A total of 449 images were included in the statistical analysis. For the seven indicators, the proportions of agreement between SW-TH01/II and sonographers' analysis results were all greater than 0.8. For the echogenicity (with very hypoechoic), aspect ratio and margin, the kappa coefficient between the two methods were above 0.75 (P < 0.001). The kappa coefficients of echogenicity (echotexture and echogenicity level), border and calcification between the two methods were above 0.6 (P < 0.001). The median time it takes for software and sonographers to interpret an image were 3 (2, 3) seconds and 26.5 (21.17, 34.33) seconds, respectively, and the difference were statistically significant (z = -18.36, P < 0.001). SW-TH01/II has a high degree of accuracy and great time efficiency benefits in judging the characteristics of thyroid nodule. It can provide more objective results and improve the efficiency of ultrasound examination. SW-TH01/II can be used to assist the sonographers in characterizing the thyroid nodule ultrasound images.

Optimizing resource allocation for Parkinson's disease (PD) motor rehabilitation necessitates identifying biomarkers of responsiveness and dynamic neuroplasticity signatures underlying efficacy. A cohort study of 52 early-stage PD patients undergoing 2-week multidisciplinary intensive rehabilitation therapy (MIRT) was conducted, which stratified participants into responders and nonresponders. A multimodal analysis of resting-state electroencephalography (EEG) microstates and functional magnetic resonance imaging (fMRI) coactivation patterns was performed to characterize MIRT-induced spatiotemporal network reorganization. Responders demonstrated clinically meaningful improvement in motor symptoms, exceeding the minimal clinically important difference threshold of 3.25 on the Unified PD Rating Scale part III, alongside significant reductions in bradykinesia and a significant enhancement in quality-of-life scores at the 3-month follow-up. Resting-state EEG in responders showed a significant attenuation in microstate C and a significant enhancement in microstate D occurrences, along with significantly increased transitions from microstate A/B to D, which significantly correlated with motor function, especially in bradykinesia gains. Concurrently, fMRI analyses identified a prolonged dwell time of the dorsal attention network coactivation/ventral attention network deactivation pattern, which was significantly inversely associated with microstate C occurrence and significantly linked to motor improvement. The identified brain spatiotemporal neural markers were validated using machine learning models to assess the efficacy of MIRT in motor rehabilitation for PD patients, achieving an average accuracy rate of 86%. These findings suggest that MIRT may facilitate a shift in neural networks from sensory processing to higher-order cognitive control, with the dynamic reallocation of attentional resources. This preliminary study validates the necessity of integrating cognitive-motor strategies for the motor rehabilitation of PD and identifies novel neural markers for assessing treatment efficacy.

The locus coeruleus (LC) is linked to the development and pathophysiology of neurodegenerative diseases such as Alzheimer's disease (AD). Magnetic resonance imaging-based LC features have shown potential to assess LC integrity in vivo. We present a deep learning-based LC segmentation and feature extraction method called Ensemble-based Locus Coeruleus Segmentation Network (ELSI-Net) and apply it to healthy aging and AD dementia datasets. Agreement to expert raters and previously published LC atlases were assessed. We aimed to reproduce previously reported differences in LC integrity in aging and AD dementia and correlate extracted features to cerebrospinal fluid (CSF) biomarkers of AD pathology. ELSI-Net demonstrated high agreement to expert raters and published atlases. Previously reported group differences in LC integrity were detected and correlations to CSF biomarkers were found. Although we found excellent performance, further evaluations on more diverse datasets from clinical cohorts are required for a conclusive assessment of ELSI-Net's general applicability. We provide a thorough evaluation of a fully automatic locus coeruleus (LC) segmentation method termed Ensemble-based Locus Coeruleus Segmentation Network (ELSI-Net) in aging and Alzheimer's disease (AD) dementia.ELSI-Net outperforms previous work and shows high agreement with manual ratings and previously published LC atlases.ELSI-Net replicates previously shown LC group differences in aging and AD.ELSI-Net's LC mask volume correlates with cerebrospinal fluid biomarkers of AD pathology.

Currently, plaque segmentation in Optical Coherence Tomography (OCT) images of coronary arteries is primarily carried out manually by physicians, and the accuracy of existing automatic segmentation techniques needs further improvement. To furnish efficient and precise decision support, automated detection of plaques in coronary OCT images holds paramount importance. For addressing these challenges, we propose a novel deep learning algorithm featuring Dense Atrous Convolution (DAC) and attention mechanism to realize high-precision segmentation and classification of Coronary artery plaques. Then, a relatively well-established dataset covering 760 original images, expanded to 8,000 using data enhancement. This dataset serves as a significant resource for future research endeavors. The experimental results demonstrate that the dice coefficients of calcified, fibrous, and lipid plaques are 0.913, 0.900, and 0.879, respectively, surpassing those generated by five other conventional medical image segmentation networks. These outcomes strongly attest to the effectiveness and superiority of our proposed algorithm in the task of automatic coronary artery plaque segmentation.

The rapid advancement of the civil aviation industry has attracted significant attention to research on pilots. However, the brain changes experienced by flight cadets following their training remain, to some extent, an unexplored territory compared to those of the general population. The aim of this study was to examine the impact of flight training on brain function by employing machine learning(ML) techniques. We collected resting-state functional magnetic resonance imaging (resting-state fMRI) data from 79 flight cadets and ground program cadets, extracting blood oxygenation level-dependent (BOLD) signal, amplitude of low frequency fluctuation (ALFF), regional homogeneity (ReHo), and functional connectivity (FC) metrics as feature inputs for ML models. After conducting feature selection using a two-sample t-test, we established various ML classification models, including Extreme Gradient Boosting (XGBoost), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Gaussian Naive Bayes (GNB). Comparative analysis of the model results revealed that the LR classifier based on BOLD signals could accurately distinguish flight cadets from the general population, achieving an AUC of 83.75% and an accuracy of 0.93. Furthermore, an analysis of the features contributing significantly to the ML classification models indicated that these features were predominantly located in brain regions associated with auditory-visual processing, motor function, emotional regulation, and cognition, primarily within the Default Mode Network (DMN), Visual Network (VN), and SomatoMotor Network (SMN). These findings suggest that flight-trained cadets may exhibit enhanced functional dynamics and cognitive flexibility.