Background: Sepsis commonly causes acute respiratory distress syndrome (ARDS), and ARDS contributes to poor prognosis in sepsis patients. Early prediction of ARDS for sepsis patients remains a clinical challenge. This study aims to develop and validate chest computed tomography (CT) radiomic-based signatures for early prediction of ARDS and assessment of individual severity in sepsis patients. Methods: In this ambispective observational cohort study, a deep learning model, a sepsis-induced acute respiratory distress syndrome (SI-ARDS) prediction neural network, will be developed to extract radiomics features of chest CT from sepsis patients. The datasets will be collected from these retrospective and prospective cohorts, including 400 patients diagnosed with sepsis-3 definition during a period from 1 May 2015 to 30 May 2022. 160 patients of the retrospective cohort will be selected as a discovering group to reconstruct the model and 40 patients of the retrospective cohort will be selected as a testing group for internal validation. Additionally, 200 patients of the prospective cohort from two hospitals will be selected as a validating group for external validation. Data pertaining to chest CT, clinical information, immune-associated inflammatory indicators and follow-up will be collected. The primary outcome is to develop and validate the model, predicting in-hospital incidence of SI-ARDS. Finally, model performance will be evaluated using the area under the curve (AUC) of receiver operating characteristic (ROC), sensitivity and specificity, using internal and external validations. Discussion: Present studies reveal that early identification and classification of the SI-ARDS is essential to improve prognosis and disease management. Chest CT has been sought as a useful diagnostic tool to identify ARDS. However, when characteristic imaging findings were clearly presented, delays in diagnosis and treatment were impossible to avoid. In this ambispective cohort study, we hope to develop a novel model incorporating radiomic signatures and clinical signatures to provide an easy-to-use and individualized prediction of SI-ARDS occurrence and severe degree in patients at early stage.

Peripheral blood smear morphological analysis remains fundamental in hematological disease diagnosis. However, existing publicly available medical datasets lack annotations for schistocytes—pivotal biomarkers integral for disease diagnosis, severity stratification, and monitoring therapeutic responsiveness. To address this gap, we constructed the first standardized blood smear dataset containing six cell types: erythrocytes, platelets, granulocytes, lymphocytes, monocytes, and schistocytes. This dataset comprises 6,150 high-resolution, pathologist-validated images with data augmentation, establishing a benchmark for complex morphological cell detection. To overcome multi-scale detection challenges (ranging from 2 to 50 μm), we propose a hierarchical multi-scale neural network framework (MCS-Net) integrating three innovations: (1) A multi-scale convolution and attention fusion module to extract local details and global semantic features; (2) A Cross-Scale Dynamic interaction detection head to mitigate high-level feature suppression of micro-targets while maintaining adequate receptive fields for large targets, enabling adaptive feature fusion in dense cell clusters; (3) A single-head self-attention mechanism to optimize leukocyte nuclear-cytoplasmic correlations, complemented by normalized Wasserstein distance for robust micro-target localization. Experiments demonstrate our framework achieves a mean average precision at 0.5 intersection-over-union (mean Average Precision (mAP)@0.5) of 0.958 ± 0.0022 on our internal dataset, surpassing the baseline by 3.9 ± 0.39%, with schistocyte detection precision reaching 90.1 ± 0.39%. The lightweight combination of the Cross-Scale Dynamic interaction detection head and normalized Wasserstein distance maintains a mAP@0.5 of 0.946 at 102 frames per second, providing an efficient solution for real-time medical imaging analysis. This work advances early screening tools for critical conditions such as Disseminated Intravascular Coagulation and Microangiopathic Hemolytic Anemia (DIC/MAHA) and drives algorithmic innovation in blood cell subtype detection.

Triage is of great importance in the procedure of the emergency diagnosis, which needs artificial intelligence to assist due to the scarcity of medical resources. However, AI models for aiding triage have difficulty in identifying levels that are difficult or ambiguous for human clinicians to distinguish. Here, to address this challenge and improve the performance of triage models, we propose KUTS, a foundational classification model for emergency triage, which leverages a knowledge prompt-tuning encoder and an uncertainty-based classifier. KUTS takes tabular data and chief complaints as input, and then generates the recommended triage level for human clinicians. On the most difficult level for human to distinguish (Level 3), our KUTS significantly outperformed the previous shallow single-modal methods, deep single-modal methods and deep multi-modal methods by an average of 14.85%, 28.38% and 11.14%, respectively. Besides, on all the triage levels, our KUTS also outperformed the previous shallow single-modal methods, deep single-modal methods and deep multi-modal methods by an average of 9.02%, 9.45% and 4.18%, respectively. KUTS provides a foundational framework and paradigm for the study of emergency triage, and can be extended to scenarios that require pre-hospital emergency treatment such as disasters and accidents.