Timely and precise identification of acute ischemic stroke (AIS) within 4.5 h is imperative for effective treatment decision-making. This study aims to construct a novel network that utilizes limited datasets to recognize AIS patients within this critical window. We conducted a retrospective analysis of 265 AIS patients who underwent both fluid attenuation inversion recovery (FLAIR) and diffusion-weighted imaging (DWI) within 24 h of symptom onset. Patients were categorized based on the time since stroke onset (TSS) into two groups: TSS ≤ 4.5 h and TSS > 4.5 h. The TSS was calculated as the time from stroke onset to MRI completion. We proposed a synchronized dual-stage network (SDS-Net) and a sequential dual-stage network (Dual-stage Net), which were comprised of infarct voxel identification and TSS classification stages. The models were trained on 181 patients and validated on an independent external cohort of 84 patients using metrics of area under the curve (AUC), sensitivity, specificity, and accuracy. A DeLong test was used to statistically compare the performance of the two models. SDS-Net achieved an accuracy of 0.844 with an AUC of 0.914 in the validation dataset, outperforming the Dual-stage Net, which had an accuracy of 0.822 and an AUC of 0.846. In the external test dataset, SDS-Net further demonstrated superior performance with an accuracy of 0.800 and an AUC of 0.879, compared to the accuracy of 0.694 and AUC of 0.744 of Dual-stage Net (P = 0.049). SDS-Net is a robust and reliable tool for identifying AIS patients within a 4.5-h treatment window using MRI. This model can assist clinicians in making timely treatment decisions, potentially improving patient outcomes.

To address the unmet need for a widely available examination for mortality prediction, this study developed a foundation visual artificial intelligence (VAI) to enhance mortality risk stratification using chest X-rays (CXRs). The VAI employed deep learning to extract CXR features and a Cox proportional hazard model to generate a hazard score ("CXR-risk"). We retrospectively collected CXRs from patients visited outpatient department and physical examination center. Subsequently, we reviewed mortality and morbidity outcomes from electronic medical records. The dataset consisted of 41,945, 10,492, 31,707, and 4441 patients in the training, validation, internal test, and external test sets, respectively. During the median follow-up of 3.2 (IQR, 1.2-6.1) years of both internal and external test sets, the "CXR-risk" demonstrated C-indexes of 0.859 (95% confidence interval (CI), 0.851-0.867) and 0.870 (95% CI, 0.844-0.896), respectively. Patients with high "CXR-risk," above 85th percentile, had a significantly higher risk of mortality than those with low risk, below 50th percentile. The addition of clinical and laboratory data and radiographic report further improved the predictive accuracy, resulting in C-indexes of 0.888 and 0.900. The VAI can provide accurate predictions of mortality and morbidity outcomes using just a single CXR, and it can complement other risk prediction indicators to assist physicians in assessing patient risk more effectively.

This study aims to assess the effectiveness of integrating Segment Anything Model (SAM) and its variant MedSAM into the automated mining, object detection, and segmentation (MODS) methodology for developing robust lung cancer detection and segmentation models without post hoc labeling of training images. In a retrospective analysis, 10,000 chest computed tomography scans from patients with lung cancer were mined. Line measurement annotations were converted to bounding boxes, excluding boxes < 1 cm or > 7 cm. The You Only Look Once object detection architecture was used for teacher-student learning to label unannotated lesions on the training images. Subsequently, a final tumor detection model was trained and employed with SAM and MedSAM for tumor segmentation. Model performance was assessed on a manually annotated test dataset, with additional evaluations conducted on an external lung cancer dataset before and after detection model fine-tuning. Bootstrap resampling was used to calculate 95% confidence intervals. Data mining yielded 10,789 line annotations, resulting in 5403 training boxes. The baseline detection model achieved an internal F1 score of 0.847, improving to 0.860 after self-labeling. Tumor segmentation using the final detection model attained internal Dice similarity coefficients (DSCs) of 0.842 (SAM) and 0.822 (MedSAM). After fine-tuning, external validation showed an F1 of 0.832 and DSCs of 0.802 (SAM) and 0.804 (MedSAM). Integrating foundational segmentation models into the MODS framework results in high-performing lung cancer detection and segmentation models using only mined clinical data. Both SAM and MedSAM hold promise as foundational segmentation models for radiology images.

Lung adenocarcinoma and squamous cell carcinoma are the two most common pathological lung cancer subtypes. Accurate diagnosis and pathological subtyping are crucial for lung cancer treatment. Solitary solid lung nodules with lobulation and spiculation signs are often indicative of lung cancer; however, in some cases, postoperative pathology finds benign solid lung nodules. It is critical to accurately identify solid lung nodules with lobulation and spiculation signs before surgery; however, traditional diagnostic imaging is prone to misdiagnosis, and studies on artificial intelligence-assisted diagnosis are few. Therefore, we introduce a volumetric SWIN Transformer-based method. It is a multi-scale, multi-task, and highly interpretable model for distinguishing between benign solid lung nodules with lobulation and spiculation signs, lung adenocarcinomas, and lung squamous cell carcinoma. The technique's effectiveness was improved by using 3-dimensional (3D) computed tomography (CT) images instead of conventional 2-dimensional (2D) images to combine as much information as possible. The model was trained using 352 of the 441 CT image sequences and validated using the rest. The experimental results showed that our model could accurately differentiate between benign lung nodules with lobulation and spiculation signs, lung adenocarcinoma, and squamous cell carcinoma. On the test set, our model achieves an accuracy of 0.9888, precision of 0.9892, recall of 0.9888, and an F1-score of 0.9888, along with a class activation mapping (CAM) visualization of the 3D model. Consequently, our method could be used as a preoperative tool to assist in diagnosing solitary solid lung nodules with lobulation and spiculation signs accurately and provide a theoretical basis for developing appropriate clinical diagnosis and treatment plans for the patients.

The parotid glands are the largest of the major salivary glands. They can harbour both benign and malignant tumours. Preoperative work-up relies on MR images and fine needle aspiration biopsy, but these diagnostic tools have low sensitivity and specificity, often leading to surgery for diagnostic purposes. The aim of this paper is (1) to develop a machine learning algorithm based on MR images characteristics to automatically classify parotid gland tumours and (2) compare its results with the diagnoses of junior and senior radiologists in order to evaluate its utility in routine practice. While automatic algorithms applied to parotid tumours classification have been developed in the past, we believe that our study is one of the first to leverage four different MRI sequences and propose a comparison with clinicians. In this study, we leverage data coming from a cohort of 134 patients treated for benign or malignant parotid tumours. Using radiomics extracted from the MR images of the gland, we train a random forest and a logistic regression to predict the corresponding histopathological subtypes. On the test set, the best results are given by the random forest: we obtain a 0.720 accuracy, a 0.860 specificity, and a 0.720 sensitivity over all histopathological subtypes, with an average AUC of 0.838. When considering the discrimination between benign and malignant tumours, the algorithm results in a 0.760 accuracy and a 0.769 AUC, both on test set. Moreover, the clinical experiment shows that our model helps to improve diagnostic abilities of junior radiologists as their sensitivity and accuracy raised by 6 % when using our proposed method. This algorithm may be useful for training of physicians. Radiomics with a machine learning algorithm may help improve discrimination between benign and malignant parotid tumours, decreasing the need for diagnostic surgery. Further studies are warranted to validate our algorithm for routine use.

Trustworthiness is crucial for artificial intelligence (AI) models in clinical settings, and a fundamental aspect of trustworthy AI is uncertainty quantification (UQ). Conformal prediction as a robust uncertainty quantification (UQ) framework has been receiving increasing attention as a valuable tool in improving model trustworthiness. An area of active research is the method of non-conformity score calculation for conformal prediction. We propose deep conformal supervision (DCS), which leverages the intermediate outputs of deep supervision for non-conformity score calculation, via weighted averaging based on the inverse of mean calibration error for each stage. We benchmarked our method on two publicly available datasets focused on medical image classification: a pneumonia chest radiography dataset and a preprocessed version of the 2019 RSNA Intracranial Hemorrhage dataset. Our method achieved mean coverage errors of 16e-4 (CI: 1e-4, 41e-4) and 5e-4 (CI: 1e-4, 10e-4) compared to baseline mean coverage errors of 28e-4 (CI: 2e-4, 64e-4) and 21e-4 (CI: 8e-4, 3e-4) on the two datasets, respectively (p < 0.001 on both datasets). Based on our findings, the baseline results of conformal prediction already exhibit small coverage errors. However, our method shows a significant improvement on coverage error, particularly noticeable in scenarios involving smaller datasets or when considering smaller acceptable error levels, which are crucial in developing UQ frameworks for healthcare AI applications.

The aim of this study is to investigate the role of [¹⁸F]-PSMA-1007 PET in differentiating high- and low-risk prostate cancer (PCa) through a robust radiomics ensemble model. This retrospective study included 143 PCa patients who underwent [¹⁸F]-PSMA-1007 PET/CT imaging. PCa areas were manually contoured on PET images and 1781 image biomarker standardization initiative (IBSI)-compliant radiomics features were extracted. A 30 times iterated preliminary analysis pipeline, comprising of the least absolute shrinkage and selection operator (LASSO) for feature selection and fivefold cross-validation for model optimization, was adopted to identify the most robust features to dataset variations, select candidate models for ensemble modelling, and optimize hyperparameters. Thirteen subsets of selected features, 11 generated from the preliminary analysis plus two additional subsets, the first based on the combination of robust and fine-tuning features, and the second only on fine-tuning features were used to train the model ensemble. Accuracy, area under curve (AUC), sensitivity, specificity, precision, and f-score values were calculated to provide models' performance. Friedman test, followed by post hoc tests corrected with Dunn-Sidak correction for multiple comparisons, was used to verify if statistically significant differences were found in the different ensemble models over the 30 iterations. The model ensemble trained with the combination of robust and fine-tuning features obtained the highest average accuracy (79.52%), AUC (85.75%), specificity (84.29%), precision (82.85%), and f-score (78.26%). Statistically significant differences (p < 0.05) were found for some performance metrics. These findings support the role of [¹⁸F]-PSMA-1007 PET radiomics in improving risk stratification for PCa, by reducing dependence on biopsies.

Medical image classification using convolutional neural networks (CNNs) is promising but often requires extensive manual tuning for optimal model definition. Neural architecture search (NAS) automates this process, reducing human intervention significantly. This study applies NAS to [18F]-Florbetaben PET cardiac images for classifying cardiac amyloidosis (CA) sub-types (amyloid light chain (AL) and transthyretin amyloid (ATTR)) and controls. Following data preprocessing and augmentation, an evolutionary cell-based NAS approach with a fixed network macro-structure is employed, automatically deriving cells' micro-structure. The algorithm is executed five times, evaluating 100 mutating architectures per run on an augmented dataset of 4048 images (originally 597), totaling 5000 architectures evaluated. The best network (NAS-Net) achieves 76.95% overall accuracy. K-fold analysis yields mean ± SD percentages of sensitivity, specificity, and accuracy on the test dataset: AL subjects (98.7 ± 2.9, 99.3 ± 1.1, 99.7 ± 0.7), ATTR-CA subjects (93.3 ± 7.8, 78.0 ± 2.9, 70.9 ± 3.7), and controls (35.8 ± 14.6, 77.1 ± 2.0, 96.7 ± 4.4). NAS-derived network performance rivals manually determined networks in the literature while using fewer parameters, validating its automatic approach's efficacy.

The differentiation of benign and malignant parotid gland tumors is of major significance as it directly affects the treatment process. In addition, it is also a vital task in terms of early and accurate diagnosis of parotid gland tumors and the determination of treatment planning accordingly. As in other diseases, the differentiation of tumor types involves several challenging, time-consuming, and laborious processes. In the study, Magnetic Resonance (MR) images of 114 patients with parotid gland tumors are used for training and testing purposes by Image Fusion (IF). After the Apparent Diffusion Coefficient (ADC), Contrast-enhanced T1-w (T1C-w), and T2-w sequences are cropped, IF (ADC, T1C-w), IF (ADC, T2-w), IF (T1C-w, T2-w), and IF (ADC, T1C-w, T2-w) datasets are obtained for different combinations of these sequences using a two-dimensional Discrete Wavelet Transform (DWT)-based fusion technique. For each of these four datasets, ResNet18, GoogLeNet, and DenseNet-201 architectures are trained separately, and thus, 12 models are obtained in the study. A Graphical User Interface (GUI) application that contains the most successful of these trained architectures for each data is also designed to support the users. The designed GUI application not only allows the fusing of different sequence images but also predicts whether the label of the fused image is benign or malignant. The results show that the DenseNet-201 models for IF (ADC, T1C-w), IF (ADC, T2-w), and IF (ADC, T1C-w, T2-w) are better than the others, with accuracies of 95.45%, 95.96%, and 92.93%, respectively. It is also noted in the study that the most successful model for IF (T1C-w, T2-w) is ResNet18, and its accuracy is equal to 94.95%.

Natural language processing (NLP) is crucial to extract information accurately from unstructured text to provide insights for clinical decision-making, quality improvement, and medical research. This study compared the performance of a rule-based NLP system and a medical-domain transformer-based model to detect negated concepts in radiology reports. Using a corpus of 984 de-identified radiology reports from a large U.S.-based academic health system (1000 consecutive reports, excluding 16 duplicates), the investigators compared the rule-based medspaCy system and the Clinical Assertion and Negation Classification Bidirectional Encoder Representations from Transformers (CAN-BERT) system to detect negated expressions of terms from RadLex, the Unified Medical Language System Metathesaurus, and the Radiology Gamuts Ontology. Power analysis determined a sample size of 382 terms to achieve α = 0.05 and β = 0.8 for McNemar's test; based on an estimate of 15% negated terms, 2800 randomly selected terms were annotated manually as negated or not negated. Precision, recall, and F1 of the two models were compared using McNemar's test. Of the 2800 terms, 387 (13.8%) were negated. For negation detection, medspaCy attained a recall of 0.795, precision of 0.356, and F1 of 0.492. CAN-BERT achieved a recall of 0.785, precision of 0.768, and F1 of 0.777. Although recall was not significantly different, CAN-BERT had significantly better precision (χ2 = 304.64; p < 0.001). The transformer-based CAN-BERT model detected negated terms in radiology reports with high precision and recall; its precision significantly exceeded that of the rule-based medspaCy system. Use of this system will improve data extraction from textual reports to support information retrieval, AI model training, and discovery of causal relationships.