To develop an interpretable machine learning (ML) model using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) radiomic data, dosimetric parameters, and clinical data for predicting radiation-induced hepatic toxicity (RIHT) in patients with hepatocellular carcinoma (HCC) following intensity-modulated radiation therapy (IMRT). A retrospective analysis of 150 HCC patients was performed, with a 7:3 ratio used to divide the data into training and validation cohorts. Radiomic features from the original MRI sequences and Delta-radiomic features were extracted. Seven ML models based on radiomics were developed: logistic regression (LR), random forest (RF), support vector machine (SVM), eXtreme Gradient Boosting (XGBoost), adaptive boosting (AdaBoost), decision tree (DT), and artificial neural network (ANN). The predictive performance of the models was evaluated using receiver operating characteristic (ROC) curve analysis and calibration curves. Shapley additive explanations (SHAP) were employed to interpret the contribution of each variable and its risk threshold. Original radiomic features and Delta-radiomic features were extracted from DCE-MRI images and filtered to generate Radiomics-scores and Delta-Radiomics-scores. These were then combined with independent risk factors (Body Mass Index (BMI), V5, and pre-Child-Pugh score(pre-CP)) identified through univariate and multivariate logistic regression and Spearman correlation analysis to construct the ML models. In the training cohort, the AUC values were 0.8651 for LR, 0.7004 for RF, 0.6349 for SVM, 0.6706 for XGBoost, 0.7341 for AdaBoost, 0.6806 for Decision Tree, and 0.6786 for ANN. The corresponding accuracies were 84.4%, 65.6%, 75.0%, 65.6%, 71.9%, 68.8%, and 71.9%, respectively. The validation cohort further confirmed the superiority of the LR model, which was selected as the optimal model. SHAP analysis revealed that Delta-radiomics made a substantial positive contribution to the model. The interpretable ML model based on radiomics provides a non-invasive tool for predicting RIHT in patients with HCC, demonstrating satisfactory discriminative performance.

This study investigated perspectives and experiences of artificial intelligence (AI) developers, clinicians and patients about the use of AI-based software in cardiac healthcare. A qualitative study took place at two hospitals in England that had trialled AI-based software use in stress echocardiography, a scan that uses ultrasound to assess heart function. Semi-structured interviews were conducted with: patients (<i>n = </i>9), clinicians (<i>n = </i>16) and AI software developers (<i>n = </i>5). Data were analysed using thematic analysis. Potential benefits identified were increasing consistency and reliability through reducing human error, and greater efficiency. Concerns included over-reliance on the AI technology, and data security. Participants discussed the need for human input and empathy within healthcare, transparency about AI use, and issues around trusting AI. Participants considered AI's role as assisting diagnosis but not replacing clinician involvement. Clinicians and patients emphasised holistic diagnosis that involves more than the scan. Clinicians considered their diagnostic ability as superior and discrepancies were managed in line with clinicians' diagnoses rather than AI reports. The practicalities of using the AI software concerned image acquisition to meet AI processing requirements and workflow integration. There was positivity towards AI use, but the AI software was considered an adjunct to clinicians rather than replacing their input. Clinicians' experiences were that their diagnostic ability remained superior to the AI, and acquiring images acceptable to AI was sometimes problematic. Despite hopes for increased efficiency through AI use, clinicians struggled to identify fit with clinical workflow to bring benefit.

Currently, there are inconsistencies among different studies on preoperative prediction of Cytokeratin 19 (CK19) expression in HCC using traditional imaging, radiomics, and deep learning. We aimed to systematically analyze and compare the performance of non-invasive methods for predicting CK19-positive HCC, thereby providing insights for the stratified management of HCC patients. A comprehensive literature search was conducted in PubMed, EMBASE, Web of Science, and the Cochrane Library from inception to February 2025. Two investigators independently screened and extracted data based on inclusion and exclusion criteria. Eligible studies were included, and key findings were summarized in tables to provide a clear overview. Ultimately, 22 studies involving 3395 HCC patients were included. 72.7% (16/22) focused on traditional imaging, 36.4% (8/22) on radiomics, 9.1% (2/22) on deep learning, and 54.5% (12/22) on combined models. The magnetic resonance imaging was the most commonly used imaging modality (19/22), and over half of the studies (12/22) were published between 2022 and 2025. Moreover, 27.3% (6/22) were multicenter studies, 36.4% (8/22) included a validation set, and only 13.6% (3/22) were prospective. The area under the curve (AUC) range of using clinical and traditional imaging was 0.560 to 0.917. The AUC ranges of radiomics were 0.648 to 0.951, and the AUC ranges of deep learning were 0.718 to 0.820. Notably, the AUC ranges of combined models of clinical, imaging, radiomics and deep learning were 0.614 to 0.995. Nevertheless, the multicenter external data were limited, with only 13.6% (3/22) incorporating validation. The combined model integrating traditional imaging, radiomics and deep learning achieves excellent potential and performance for predicting CK19 in HCC. Based on current limitations, future research should focus on building an easy-to-use dynamic online tool, combining multicenter-multimodal imaging and advanced deep learning approaches to enhance the accuracy and robustness of model predictions.

Lung cancer is a major cause of cancer-related deaths, and early diagnosis and treatment are crucial for improving patients' survival outcomes. In this paper, we propose to employ convolutional neural networks to model the non-linear relationship between the risk of lung cancer and the lungs' morphology revealed in the CT images. We apply a mini-batched loss that extends the Cox proportional hazards model to handle the non-convexity induced by neural networks, which also enables the training of large data sets. Additionally, we propose to combine mini-batched loss and binary cross-entropy to predict both lung cancer occurrence and the risk of mortality. Simulation results demonstrate the effectiveness of both the mini-batched loss with and without the censoring mechanism, as well as its combination with binary cross-entropy. We evaluate our approach on the National Lung Screening Trial data set with several 3D convolutional neural network architectures, achieving high AUC and C-index scores for lung cancer classification and survival prediction. These results, obtained from simulations and real data experiments, highlight the potential of our approach to improving the diagnosis and treatment of lung cancer.

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.

Increasing evidence suggests that non-operative management (NOM) with antibiotics could serve as a safe alternative to surgery for the treatment of uncomplicated acute appendicitis (AA). However, accurately differentiating between uncomplicated and complicated AA remains challenging. Our aim was to develop and validate machine-learning-based diagnostic models to differentiate uncomplicated from complicated AA. This was a multicenter cohort trial conducted from January 2021 and December 2022 across five tertiary hospitals. Three distinct diagnostic models were created, namely, the clinical-parameter-based model, the CT-radiomics-based model, and the clinical-radiomics-fused model. These models were developed using a comprehensive set of eight machine-learning algorithms, which included logistic regression (LR), support vector machine (SVM), random forest (RF), decision tree (DT), gradient boosting (GB), K-nearest neighbors (KNN), Gaussian Naïve Bayes (GNB), and multi-layer perceptron (MLP). The performance and accuracy of these diverse models were compared. All models exhibited excellent diagnostic performance in the training cohort, achieving a maximal AUC of 1.00. For the clinical-parameter model, the GB classifier yielded the optimal AUC of 0.77 (95% confidence interval [CI]: 0.64-0.90) in the testing cohort, while the LR classifier yielded the optimal AUC of 0.76 (95% CI: 0.66-0.86) in the validation cohort. For the CT-radiomics-based model, GB classifier achieved the best AUC of 0.74 (95% CI: 0.60-0.88) in the testing cohort, and SVM yielded an optimal AUC of 0.63 (95% CI: 0.51-0.75) in the validation cohort. For the clinical-radiomics-fused model, RF classifier yielded an optimal AUC of 0.84 (95% CI: 0.74-0.95) in the testing cohort and 0.76 (95% CI: 0.67-0.86) in the validation cohort. An open-access, user-friendly online tool was developed for clinical application. This multicenter study suggests that the clinical-radiomics-fused model, constructed using RF algorithm, effectively differentiated between complicated and uncomplicated AA.

Approaches studying the dynamics of resting-state functional magnetic resonance imaging (rs-fMRI) activity often focus on time-resolved functional connectivity (tr-FC). While many tr-FC approaches have been proposed, most are linear approaches, e.g. computing the linear correlation at a timestep or within a window. In this work, we propose to use a generative non-linear deep learning model, a disentangled variational autoencoder (DSVAE), that factorizes out window-specific (context) information from timestep-specific (local) information. This has the advantage of allowing our model to capture differences at multiple temporal scales. We find that by separating out temporal scales our model's window-specific embeddings, or as we refer to them, context embeddings, more accurately separate windows from schizophrenia patients and control subjects than baseline models and the standard tr-FC approach in a low-dimensional space. Moreover, we find that for individuals with schizophrenia, our model's context embedding space is significantly correlated with both age and symptom severity. Interestingly, patients appear to spend more time in three clusters, one closer to controls which shows increased visual-sensorimotor, cerebellar-subcortical, and reduced cerebellar-visual functional network connectivity (FNC), an intermediate station showing increased subcortical-sensorimotor FNC, and one that shows decreased visual-sensorimotor, decreased subcortical-sensorimotor, and increased visual-subcortical domains. We verify that our model captures features that are complementary to - but not the same as - standard tr-FC features. Our model can thus help broaden the neuroimaging toolset in analyzing fMRI dynamics and shows potential as an approach for finding psychiatric links that are more sensitive to individual and group characteristics.

This study evaluates the impact of artificial intelligence (AI) assistance on the diagnostic performance of radiologists with varying levels of experience in interpreting mammograms in a Malaysian tertiary referral center, particularly in women with dense breasts. A retrospective study including 434 digital mammograms interpreted by two general radiologists (12 and 6 years of experience) and two trainees (2 years of experience). Diagnostic performance was assessed with and without AI assistance (Lunit INSIGHT MMG), using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC). Inter-reader agreement was measured using kappa statistics. AI assistance significantly improved the diagnostic performance of all reader groups across all metrics (p < 0.05). The senior radiologist consistently achieved the highest sensitivity (86.5% without AI, 88.0% with AI) and specificity (60.5% without AI, 59.2% with AI). The junior radiologist demonstrated the highest PPV (56.9% without AI, 74.6% with AI) and NPV (90.3% without AI, 92.2% with AI). The trainees showed the lowest performance, but AI significantly enhanced their accuracy. AI assistance was particularly beneficial in interpreting mammograms of women with dense breasts. AI assistance significantly enhances the diagnostic accuracy and consistency of radiologists in mammogram interpretation, with notable benefits for less experienced readers. These findings support the integration of AI into clinical practice, particularly in resource-limited settings where access to specialized breast radiologists is constrained.

Machine learning (ML) algorithms play a crucial role in the early and accurate diagnosis of Alzheimer's Disease (AD), which is essential for effective treatment planning. However, existing methods are not well-suited for identifying Mild Cognitive Impairment (MCI), a critical transitional stage between normal aging and AD. This inadequacy is primarily due to label imbalance and bias from different sensitve attributes in MCI classification. To overcome these challenges, we have designed an end-to-end fairness-aware approach for label-imbalanced classification, tailored specifically for neuroimaging data. This method, built on the recently developed FACIMS framework, integrates into STREAMLINE, an automated ML environment. We evaluated our approach against nine other ML algorithms and found that it achieves comparable balanced accuracy to other methods while prioritizing fairness in classifications with five different sensitive attributes. This analysis contributes to the development of equitable and reliable ML diagnostics for MCI detection.

The use of artificial intelligence (AI) in medicine has shown promise to improve the quality of healthcare decisions. However, AI can be biased in a manner that produces unfair predictions for certain demographic subgroups. In MIMIC-CXR, a publicly available dataset of over 300,000 chest X-ray images, diagnostic AI has been shown to have a higher false negative rate for racial minorities. We evaluated the capacity of synthetic data augmentation, oversampling, and demographic-based corrections to enhance the fairness of AI predictions. We show that adjusting unfair predictions for demographic attributes, such as race, is ineffective at improving fairness or predictive performance. However, using oversampling and synthetic data augmentation to modify disease prevalence reduced such disparities by 74.7% and 10.6%, respectively. Moreover, such fairness gains were accomplished without reduction in performance (95% CI AUC: [0.816, 0.820] versus [0.810, 0.819] versus [0.817, 0.821] for baseline, oversampling, and augmentation, respectively).