This study explores the application of Retrieval-Augmented Generation (RAG) for question answering in radiology, an area where intelligent systems can significantly impact clinical decision-making. A preliminary experiment tested a naive RAG setup on nice radiology-specific questions with a textbook as the reference source, showing moderate improvements over baseline methods. The paper discusses lessons learned and potential enhancements for RAG in handling radiology knowledge, suggesting pathways for future research in integrating intelligent health systems in medical practice.

Radiogenomics holds promise in identifying molecular alterations in nonsmall cell lung cancer (NSCLC) using imaging features. Previously, we developed a radiogenomics model to predict epidermal growth factor receptor (EGFR) mutations based on contrast-enhanced computed tomography (CECT) in NSCLC patients. The current study aimed to externally validate this model using a publicly available National Institutes of Health (NIH)-based NSCLC dataset and assess the effect of EGFR mutation prevalence on model performance through synthetic minority oversampling technique (SMOTE). The original radiogenomics model was validated on an independent NIH cohort (n=140). For assessing the influence of disease prevalence, six SMOTE-augmented datasets were created, simulating EGFR mutation prevalence from 25% to 50%. Seven models were developed (one from original data, six SMOTE-augmented), each undergoing rigorous cross-validation, feature selection, and logistic regression modeling. Models were tested against the NIH cohort. Performance was compared using area under the receiver operating characteristic curve (Area Under the Curve [AUC]), and differences between radiomic-only, clinical-only, and combined models were statistically assessed. External validation revealed poor diagnostic performance for both our model and a previously published EGFR radiomics model (AUC ∼0.5). The clinical model alone achieved higher diagnostic accuracy (AUC 0.74). SMOTE-augmented models showed increased sensitivity but did not improve overall AUC compared to the clinical-only model. Changing EGFR mutation prevalence had minimal impact on AUC, challenging previous assumptions about the influence of sample imbalance on model performance. External validation failed to reproduce prior radiogenomics model performance, while clinical variables alone retained strong predictive value. SMOTE-based oversampling did not improve diagnostic accuracy, suggesting that, in EGFR prediction, radiomics may offer limited value beyond clinical data. Emphasis on robust external validation and data-sharing is essential for future clinical implementation of radiogenomic models.

Medical imaging constitutes critical information in the diagnostic and prognostic evaluation of patients, as it serves to uncover a broad spectrum of pathologies and deviances. Clinical practitioners who carry out medical image screening are primarily reliant on their knowledge and experience for disease diagnosis. Convolutional Neural Networks (CNNs) hold the potential to serve as a formidable decision-support tool in the realm of medical image analysis due to their high capacity to extract hierarchical features and effectuate direct classification and segmentation from image data. However, CNNs contain a myriad of hyperparameters and optimizing these hyperparameters poses a major obstacle to the effective implementation of CNNs. In this work, a two-phase Bayesian Optimization-derived Scheduling (BOS) approach is proposed for hyperparameter optimization for the head and cancerous tissue segmentation tasks. We proposed this two-phase BOS approach to incorporate both rapid convergences in the first training phase and slower (but without overfitting) improvements in the last training phase. Furthermore, we found that batch size and learning rate have a significant impact on the training process, but optimizing them separately can lead to sub-optimal hyperparameter combinations. Therefore, batch size and learning rate have been coupled as the batch size to learning rate (B2L) ratio and utilized in the optimization process to optimize both simultaneously. The optimized hyperparameters have been tested for a three-dimensional V-Net model with computed tomography (CT) and positron emission tomography (PET) scans to segment and classify cancerous and noncancerous tissues. The results of 10-fold cross-validation indicate that the optimal batch size to learning rate (B2L) ratio for each phase of the training method can improve the overall medical image segmentation performance.

Radiotherapy plays a critical role in treating esophageal cancer, but individual responses vary significantly, impacting patient outcomes. This study integrates machine learning-driven multimodal radiomics and transcriptomics to develop predictive models for radiotherapy sensitivity and prognosis in esophageal cancer. We applied the SEResNet101 deep learning model to imaging and transcriptomic data from the UCSC Xena and TCGA databases, identifying prognosis-associated genes such as STUB1, PEX12, and HEXIM2. Using Lasso regression and Cox analysis, we constructed a prognostic risk model that accurately stratifies patients based on survival probability. Notably, STUB1, an E3 ubiquitin ligase, enhances radiotherapy sensitivity by promoting the ubiquitination and degradation of SRC, a key oncogenic protein. In vitro and in vivo experiments confirmed that STUB1 overexpression or SRC silencing significantly improves radiotherapy response in esophageal cancer models. These findings highlight the predictive power of multimodal data integration for individualized radiotherapy planning and underscore STUB1 as a promising therapeutic target for enhancing radiotherapy efficacy in esophageal cancer.

Early identification of individuals who progress from normal cognition (NC) to mild cognitive impairment (MCI) may help prevent cognitive decline. We aimed to build predictive models using radiomic features of the bilateral hippocampus in combination with scores from neuropsychological assessments. We utilized the Alzheimer's Disease Neuroimaging Initiative (ADNI) database to study 175 NC individuals, identifying 50 who progressed to MCI within seven years. Employing the Least Absolute Shrinkage and Selection Operator (LASSO) on T1-weighted images, we extracted and refined hippocampal features. Classification models, including Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and light gradient boosters (LightGBM), were built based on significant neuropsychological scores. Model validation was conducted using 5-fold cross-validation, and hyperparameters were optimized with Scikit-learn, using an 80:20 data split for training and testing. We found that the LightGBM model achieved an area under the receiver operating characteristic (ROC) curve (AUC) value of 0.89 and an accuracy of 0.79 in the training set, and an AUC value of 0.80 and an accuracy of 0.74 in the test set. The study identified that T1-weighted magnetic resonance imaging radiomic of the hippocampus would be used to predict the progression to MCI at the normal cognitive stage, which might provide a new insight into clinical research.

Integrating deep learning into clinical workflows for medical image analysis holds promise for improving diagnostic accuracy. However, strict data privacy regulations and the sensitivity of clinical IT infrastructure limit the deployment of cloud-based solutions. This paper introduces WebIPred, a web-based application that loads deep learning models directly within the client's web browser, protecting patient privacy while maintaining compatibility with clinical IT environments. WebIPred supports the application of pre-trained models published on Zenodo and other repositories, allowing clinicians to apply these models to real patient data without the need for extensive technical knowledge. This paper outlines WebIPred's model integration system, prediction workflow, and privacy features. Our results show that WebIPred offers a privacy-protecting and flexible application for image classification, only relying on client-side processing. WebIPred combines its strong commitment to data privacy and security with a user-friendly interface that makes it easy for clinicians to integrate AI into their workflows.

Recently, combining the strategy of consistency regularization with uncertainty estimation has shown promising performance on semi-supervised medical image segmentation tasks. However, most existing methods estimate the uncertainty solely based on the outputs of a single neural network, which results in imprecise uncertainty estimations and eventually degrades the segmentation performance. In this paper, we propose a novel Uncertainty Co-estimator (UnCo) framework to deal with this problem. Inspired by the co-training technique, UnCo establishes two different mean-teacher modules (i.e., two pairs of teacher and student models), and estimates three types of uncertainty from the multi-source predictions generated by these models. Through combining these uncertainties, their differences will help to filter out incorrect noise in each estimate, thus allowing the final fused uncertainty maps to be more accurate. These resulting maps are then used to enhance a cross-consistency regularization imposed between the two modules. In addition, UnCo also designs an internal consistency regularization within each module, so that the student models can aggregate diverse feature information from both modules, thus promoting the semi-supervised segmentation performance. Finally, an adversarial constraint is introduced to maintain the model diversity. Experimental results on four medical image datasets indicate that UnCo can achieve new state-of-the-art performance on both 2D and 3D semi-supervised segmentation tasks. The source code will be available at https://github.com/z1010x/UnCo.

This paper addresses the problem of detecting possible serious bacterial infection (pSBI) of infancy, i.e. a clinical presentation consistent with bacterial sepsis in newborn infants using cranial ultrasound (cUS) images. The captured image set for each patient enables multiview imagery: coronal and sagittal, with geometric overlap. To exploit this geometric relation, we develop a new learning framework, called the intersection-guided Crossview Local- and Image-level Fusion Network (CLIF-Net). Our technique employs two distinct convolutional neural network branches to extract features from coronal and sagittal images with newly developed multi-level fusion blocks. Specifically, we leverage the spatial position of these images to locate the intersecting region. We then identify and enhance the semantic features from this region across multiple levels using cross-attention modules, facilitating the acquisition of mutually beneficial and more representative features from both views. The final enhanced features from the two views are then integrated and projected through the image-level fusion layer, outputting pSBI and non-pSBI class probabilities. We contend that our method of exploiting multi-view cUS images enables a first of its kind, robust 3D representation tailored for pSBI detection. When evaluated on a dataset of 302 cUS scans from Mbale Regional Referral Hospital in Uganda, CLIF-Net demonstrates substantially enhanced performance, surpassing the prevailing state-of-the-art infection detection techniques.

Artificial intelligence (AI) algorithms can provide rapid and precise radiographic bone age (BA) assessment. This study assessed the effects of an AI algorithm on the BA assessment performance of radiologists, and evaluated how automation bias could affect radiologists. In this prospective randomized crossover study, six radiologists with varying levels of experience (senior, mi-level, and junior) assessed cases from a test set of 200 standard BA radiographs. The test set was equally divided into two subsets: datasets A and B. Each radiologist assessed BA independently without AI assistance (A- B-) and with AI assistance (A+ B+). We used the mean of assessments made by two experts as the ground truth for accuracy assessment; subsequently, we calculated the mean absolute difference (MAD) between the radiologists' BA predictions and ground-truth BA and evaluated the proportion of estimates for which the MAD exceeded one year. Additionally, we compared the radiologists' performance under conditions of early AI assistance with their performance under conditions of delayed AI assistance; the radiologists were allowed to reject AI interpretations. The overall accuracy of senior, mid-level, and junior radiologists improved significantly with AI assistance than without AI assistance (MAD: 0.74 vs. 0.46 years, p < 0.001; proportion of assessments for which MAD exceeded 1 year: 24.0% vs. 8.4%, p < 0.001). The proportion of improved BA predictions with AI assistance (16.8%) was significantly higher than that of less accurate predictions with AI assistance (2.3%; p < 0.001). No consistent timing effect was observed between conditions of early and delayed AI assistance. Most disagreements between radiologists and AI occurred over images for patients aged ≤8 years. Senior radiologists had more disagreements than other radiologists. The AI algorithm improved the BA assessment accuracy of radiologists with varying experience levels. Automation bias was prone to affect less experienced radiologists.