Correct automatic analysis of a medical report requires the identification of negations and their scopes. Since most of available training data comes from medical texts in English, it usually takes additional work to apply to non-English languages. Here, we introduce a supervised learning method for automatically identifying and determining the scopes and negation cues in French medical reports using language models based on BERT. Using a new private corpus of French-language chest CT scan reports with consistent annotation, we first fine-tuned five available transformer models on the negation cue and scope identification task. Subsequently, we extended the methodology by modifying the optimal model to encompass a wider range of clinical notes and reports (not limited to radiology reports) and more heterogeneous annotations. Lastly, we tested the generated model on its initial mask-filling task to ensure there is no catastrophic forgetting. On a corpus of thoracic CT scan reports annotated by four annotators within our team, our method reaches a F1-score of 99.4% for cue detection and 94.5% for scope detection, thus equaling or improving state-of-the art performance. On more generic biomedical reports, annotated with more heterogeneous rules, the quality of the automatic analysis of course decreases, but our best-of-the class model still delivers very good performance, with F1-scores of 98.2% (cue detection), and 90.9% (scope detection). Moreover, we show that fine-tuning the original model for the negation identification task preserves or even improves its performance on its initial fill-mask task, depending on the lemmatization. Considering the performance of our fine-tuned model for the detection of negation cues and scopes in medical reports in French and its robustness with respect to the diversity of the annotation rules and the type of biomedical data, we conclude that it is suited for use in a real-life clinical context.

This study comprehensively explores knowledge distillation frameworks for COVID-19 and lung cancer classification using chest X-ray (CXR) images. We employ high-capacity teacher models, including VGG19 and lightweight Vision Transformers (Visformer-S and AutoFormer-V2-T), to guide the training of a compact, hardware-aware student model derived from the OFA-595 supernet. Our approach leverages hybrid supervision, combining ground-truth labels with teacher models' soft targets to balance accuracy and computational efficiency. We validate our models on two benchmark datasets: COVID-QU-Ex and LCS25000, covering multiple classes, including COVID-19, healthy, non-COVID pneumonia, lung, and colon cancer. To interpret the spatial focus of the models, we employ Score-CAM-based visualizations, which provide insight into the reasoning process of both teacher and student networks. The results demonstrate that the distilled student model maintains high classification performance with significantly reduced parameters and inference time, making it an optimal choice in resource-constrained clinical environments. Our work underscores the importance of combining model efficiency with explainability for practical, trustworthy medical AI solutions.

Accurate preoperative assessment of occult lymph node metastasis (OLNM) plays a crucial role in informing therapeutic decision-making for lung cancer patients. Computed tomography (CT) is the most widely used imaging modality for preoperative work-up. The aim of this study was to develop and validate a CT-based machine learning model integrating intra-and peri-tumoral features to predict OLNM in lung cancer patients. Eligible patients with peripheral lung cancer confirmed by radical surgical excision with systematic lymphadenectomy were retrospectively recruited from January 2019 to December 2021. 1688 radiomics features were obtained from each manually segmented VOI which was composed of gross tumor volume (GTV) covering the boundary of entire tumor and three peritumoral volumes (PTV3, PTV6 and PTV9) that capture the region outside the tumor. A clinical-radiomics model incorporating radiomics signature, independent clinical factors and CT semantic features was established via multivariable logistic regression analysis and presented as a nomogram. Model performance was evaluated by discrimination, calibration, and clinical utility. Overall, 591 patients were recruited in the training cohort and 253 in the validation cohort. The radiomics signature of PTV9 showed superior diagnostic performance compared to PTV3 and PTV6 models. Integrating GPTV radiomics signature (incorporating Rad-score of GTV and PTV9) with clinical risk factor of serum CEA levels and CT imaging features of lobulation sign and tumor-pleura relationship demonstrated favorable accuracy in predicting OLNM in the training cohort (AUC, 0.819; 95% CI: 0.780-0.857) and validation cohort (AUC, 0.801; 95% CI: 0.741-0.860). The predictive performance of the clinical-radiomics model demonstrated statistically significant superiority over that of the clinical model in both cohorts (all p < 0.05). The clinical-radiomics model was able to serve as a noninvasive preoperative prediction tool for personalized risk assessment of OLNM in peripheral lung cancer patients.

Supervised learning techniques have proven their efficacy in many applications with abundant data. However, applying these methods to medical imaging is challenging due to the scarcity of data, given the high acquisition costs and intricate data characteristics of those images, thereby limiting the full potential of deep neural networks. To address the lack of data, augmentation techniques leverage geometry, color, and the synthesis ability of generative models (GMs). Despite previous efforts, gaps in the generation process limit the impact of data augmentation to improve understanding of medical images, e.g., the highly structured nature of some domains, such as X-ray images, is ignored. Current GMs rely solely on the network's capacity to blindly synthesize augmentations that preserve semantic relationships of chest X-ray images, such as anatomical restrictions, representative structures, or structural similarities consistent across datasets. In this paper, we introduce a novel GM that leverages the structural resemblance of medical images by learning a latent graph representation (LGR). We design an end-to-end model to learn (i) a LGR that captures the intrinsic structure of X-ray images and (ii) a graph convolutional network (GCN) that reconstructs the X-ray image from the LGR. We employ adversarial training to guide the generator and discriminator models in learning the distribution of the learned LGR. Using the learned GCN, our approach generates structure-preserving synthetic images by mapping generated LGRs to X-ray. Additionally, we evaluate the learned graph representation for other tasks, such as X-ray image classification and segmentation. Numerical experiments demonstrate the efficacy of our approach, increasing performance up to $3\%$ and $2\%$ for classification and segmentation, respectively.

To construct a prediction model for radiation pneumonitis (RP) in lung cancer patients based on clinical information, medical text, and 2.5D deep learning (DL) features. A total of 356 patients with lung cancer from the Heping Campus of the First Hospital of China Medical University were randomly divided at a 7:3 ratio into training and validation cohorts, and 238 patients from 3 other centers were included in the testing cohort for assessing model generalizability. We used the term frequency-inverse document frequency method to generate numerical vectors from computed tomography (CT) report texts. The CT and radiation therapy dose slices demonstrating the largest lung region of interest across the coronal and transverse planes were considered as the central slice; moreover, 3 slices above and below the central slice were selected to create comprehensive 2.5D data. We extracted DL features via DenseNet121, DenseNet201, and Twins-SVT and integrated them via multi-instance learning (MIL) fusion. The performances of the 2D and 3D DL models were also compared with the performance of the 2.5D MIL model. Finally, RP prediction models based on clinical information, medical text, and 2.5D DL features were constructed, validated, and tested. The 2.5D MIL model based on CT was significantly better than the 2D and 3D DL models in the training, validation, and test cohorts. The 2.5D MIL model based on radiation therapy dose was considered to be the optimal model in the test1 cohort, whereas the 2D model was considered to be the optimal model in the training, validation, and test3 cohorts, with the 3D model being the optimal model in the test2 cohort. A combined model achieved Area Under Curve values of 0.964, 0.877, 0.868, 0.884, and 0.849 in the training, validation, test1, test2, and test3 cohorts, respectively. We propose an RP prediction model that integrates clinical information, medical text, and 2.5D MIL features, which provides new ideas for predicting the side effects of radiation therapy.

COVID19 took the world by storm since December 2019. A highly infectious communicable disease, COVID19 is caused by the SARSCoV2 virus. By March 2020, the World Health Organization (WHO) declared COVID19 as a global pandemic. A pandemic in the 21st century after almost 100 years was something the world was not prepared for, which resulted in the deaths of around 1.6 million people worldwide. The most common symptoms of COVID19 were associated with the respiratory system and resembled a cold, flu, or pneumonia. After extensive research, doctors and scientists concluded that the main reason for lives being lost due to COVID19 was failure of the respiratory system. Patients were dying gasping for breath. Top healthcare systems of the world were failing badly as there was an acute shortage of hospital beds, oxygen cylinders, and ventilators. Many were dying without receiving any treatment at all. The aim of this project is to help doctors decide the severity of COVID19 by reading the patient's Computed Tomography (CT) scans of the lungs. Computer models are less prone to human error, and Machine Learning or Neural Network models tend to give better accuracy as training improves over time. We have decided to use a Convolutional Neural Network model. Given that a patient tests positive, our model will analyze the severity of COVID19 infection within one month of the positive test result. The severity of the infection may be promising or unfavorable (if it leads to intubation or death), based entirely on the CT scans in the dataset.

Accurate non-invasive prediction of histopathologic invasiveness and recurrence risk remains a clinical challenge in resectable non-small cell lung cancer (NSCLC). We developed and validated the Edge Proximity Score (EPS), a novel [¹⁸F]FDG PET/CT-based spatial imaging feature that quantifies the displacement of SUVmax relative to the tumor centroid and perimeter, to assess tumor aggressiveness and predict progression-free survival (PFS). This retrospective study included 244 NSCLC patients with preoperative [¹⁸F]FDG PET/CT. EPS was computed from normalized SUVmax-to-centroid and SUVmax-to-perimeter distances. A total of 115 PET radiomics features were extracted and standardized. Eight machine learning models (80:20 split) were trained to predict lymphovascular invasion (LVI), visceral pleural invasion (VPI), and spread through air spaces (STAS), with feature importance assessed using SHAP. Prognostic analysis was conducted using multivariable Cox regression. A survival prediction model incorporating EPS was externally validated in the TCIA cohort. RNA sequencing data from 76 TCIA patients were used for transcriptomic and immune profiling. EPS was significantly elevated in tumors with LVI, VPI, and STAS (P < 0.001), consistently ranked among the top SHAP features, and was an independent predictor of PFS (HR = 2.667, P = 0.015). The EPS-based nomogram achieved AUCs of 0.67, 0.70, and 0.68 for predicting 1-, 3-, and 5-year PFS in the TCIA validation cohort. High EPS was associated with proliferative and metabolic gene signatures, whereas low EPS was linked to immune activation and neutrophil infiltration. EPS is a biologically relevant, non-invasive imaging biomarker that may improve risk stratification in NSCLC.

Lung cancer is one of the leading causes of cancer-related mortality worldwide. The diagnosis of this disease remains a challenge due to the subtle and ambiguous nature of early-stage symptoms and imaging findings. Deep learning approaches, specifically Convolutional Neural Networks (CNNs), have significantly advanced medical image analysis. However, conventional architectures such as ResNet50 that rely on first-order pooling often fall short. This study aims to overcome the limitations of CNNs in lung cancer classification by proposing a novel and dynamic model named LungSE-SOP. The model is based on Second-Order Pooling (SOP) and Squeeze-and-Excitation Networks (SENet) within a ResNet50 backbone to improve feature representation and class separation. A novel Dynamic Feature Enhancement (DFE) module is also introduced, which dynamically adjusts the flow of information through SOP and SENet blocks based on learned importance scores. The model was trained using a publicly available IQ-OTH/NCCD lung cancer dataset. The performance of the model was assessed using various metrics, including the accuracy, precision, recall, F1-score, ROC curves, and confidence intervals. For multiclass tumor classification, our model achieved 98.6% accuracy for benign, 98.7% for malignant, and 99.9% for normal cases. Corresponding F1-scores were 99.2%, 99.8%, and 99.9%, respectively, reflecting the model's high precision and recall across all tumor types and its strong potential for clinical deployment.

In medical imaging, inter-observer variability among radiologists often introduces label uncertainty, particularly in modalities where visual interpretation is subjective. Lung ultrasound (LUS) is a prime example-it frequently presents a mixture of highly ambiguous regions and clearly discernible structures, making consistent annotation challenging even for experienced clinicians. In this work, we introduce a novel approach to both labeling and training AI models using expert-supplied, per-pixel confidence values. Rather than treating annotations as absolute ground truth, we design a data annotation protocol that captures the confidence that radiologists have in each labeled region, modeling the inherent aleatoric uncertainty present in real-world clinical data. We demonstrate that incorporating these confidence values during training leads to improved segmentation performance. More importantly, we show that this enhanced segmentation quality translates into better performance on downstream clinically-critical tasks-specifically, estimating S/F oxygenation ratio values, classifying S/F ratio change, and predicting 30-day patient readmission. While we empirically evaluate many methods for exposing the uncertainty to the learning model, we find that a simple approach that trains a model on binarized labels obtained with a (60%) confidence threshold works well. Importantly, high thresholds work far better than a naive approach of a 50% threshold, indicating that training on very confident pixels is far more effective. Our study systematically investigates the impact of training with varying confidence thresholds, comparing not only segmentation metrics but also downstream clinical outcomes. These results suggest that label confidence is a valuable signal that, when properly leveraged, can significantly enhance the reliability and clinical utility of AI in medical imaging.

Accurate lung cancer risk prediction remains challenging due to substantial variability across patient populations and clinical settings -- no single model performs best for all cohorts. To address this, we propose a personalized lung cancer risk prediction agent that dynamically selects the most appropriate model for each patient by combining cohort-specific knowledge with modern retrieval and reasoning techniques. Given a patient's CT scan and structured metadata -- including demographic, clinical, and nodule-level features -- the agent first performs cohort retrieval using FAISS-based similarity search across nine diverse real-world cohorts to identify the most relevant patient population from a multi-institutional database. Second, a Large Language Model (LLM) is prompted with the retrieved cohort and its associated performance metrics to recommend the optimal prediction algorithm from a pool of eight representative models, including classical linear risk models (e.g., Mayo, Brock), temporally-aware models (e.g., TDVIT, DLSTM), and multi-modal computer vision-based approaches (e.g., Liao, Sybil, DLS, DLI). This two-stage agent pipeline -- retrieval via FAISS and reasoning via LLM -- enables dynamic, cohort-aware risk prediction personalized to each patient's profile. Building on this architecture, the agent supports flexible and cohort-driven model selection across diverse clinical populations, offering a practical path toward individualized risk assessment in real-world lung cancer screening.