Large language models (LLMs) have the potential to objectively evaluate radiology resident reports; however, research on their use for feedback in radiology training and assessment of resident skill development remains limited. This study aimed to assess the effectiveness of LLMs in revising radiology reports by comparing them with reports verified by board-certified radiologists and to analyze the progression of resident's reporting skills over time. To identify the LLM that best aligned with human radiologists, 100 reports were randomly selected from 7376 reports authored by nine first-year radiology residents. The reports were evaluated based on six criteria: (1) addition of missing positive findings, (2) deletion of findings, (3) addition of negative findings, (4) correction of the expression of findings, (5) correction of the diagnosis, and (6) proposal of additional examinations or treatments. Reports were segmented into four time-based terms, and 900 reports (450 CT and 450 MRI) were randomly chosen from the initial and final terms of the residents' first year. The revised rates for each criterion were compared between the first and last terms using the Wilcoxon Signed-Rank test. Among the three LLMs-ChatGPT-4 Omni (GPT-4o), Claude-3.5 Sonnet, and Claude-3 Opus-GPT-4o demonstrated the highest level of agreement with board-certified radiologists. Significant improvements were noted in Criteria 1-3 when comparing reports from the first and last terms (Criteria 1, 2, and 3; P < 0.001, P = 0.023, and P = 0.004, respectively) using GPT-4o. No significant changes were observed for Criteria 4-6. Despite this, all criteria except for Criteria 6 showed progressive enhancement over time. LLMs can effectively provide feedback on commonly corrected areas in radiology reports, enabling residents to objectively identify and improve their weaknesses and monitor their progress. Additionally, LLMs may help reduce the workload of radiologists' mentors.

Respiratory motion during PET acquisition may produce lesion blurring. Ultra-fast 20-second breath-hold (U2BH) PET reduces respiratory motion artifacts, but the shortened scanning time increases statistical noise and may affect diagnostic quality. This study aims to denoise the U2BH PET images using a deep learning (DL)-based method. The study was conducted on two datasets collected from five scanners where the first dataset included 1272 retrospectively collected full-time PET data while the second dataset contained 46 prospectively collected U2BH and the corresponding full-time PET/CT images. A robust and data-efficient DL method called mask vision transformer (Mask-ViT) was proposed which, after fine-tuned on a limited number of training data from a target scanner, was directly applied to unseen testing data from new scanners. The performance of Mask-ViT was compared with state-of-the-art DL methods including U-Net and C-Gan taking the full-time PET images as the reference. Statistical analysis on image quality metrics were carried out with Wilcoxon signed-rank test. For clinical evaluation, two readers scored image quality on a 5-point scale (5 = excellent) and provided a binary assessment for diagnostic quality evaluation. The U2BH PET images denoised by Mask-ViT showed statistically significant improvement over U-Net and C-Gan on image quality metrics (p < 0.05). For clinical evaluation, Mask-ViT exhibited a lesion detection accuracy of 91.3%, 90.4% and 91.7%, when it was evaluated on three different scanners. Mask-ViT can effectively enhance the quality of the U2BH PET images in a data-efficient generalization setup. The denoised images meet clinical diagnostic requirements of lesion detectability.

Attention Mechanism (AM) selectively focuses on essential information for imaging tasks and captures relationships between regions from distant pixel neighborhoods to compute feature representations. Accelerated magnetic resonance image (MRI) reconstruction can benefit from AM, as the imaging process involves acquiring Fourier domain measurements that influence the image representation in a non-local manner. However, AM-based models are more adept at capturing low-frequency information and have limited capacity in constructing high-frequency representations, restricting the models to smooth reconstruction. Secondly, AM-based models need mode-specific retraining for multimodal MRI data as their knowledge is restricted to local contextual variations within modes that might be inadequate to capture the diverse transferable features across heterogeneous data domains. To address these challenges, we propose a neuromodulation-based discriminative multi-spectral AM for scalable MRI reconstruction, that can (i) propagate the context-aware high-frequency details for high-quality image reconstruction, and (ii) capture features reusable to deviated unseen domains in multimodal MRI, to offer high practical value for the healthcare industry and researchers. The proposed network consists of a spectral filtering convolutional neural network to capture mode-specific transferable features to generalize to deviated MRI data domains and a dynamic high-pass kernel generation transformer that focuses on high-frequency details for improved reconstruction. We have evaluated our model on various aspects, such as comparative studies in supervised and self-supervised learning, diffusion model-based training, closed-set and open-set generalization under heterogeneous MRI data, and interpretation-based analysis. Our results show that the proposed method offers scalable and high-quality reconstruction with best improvement margins of ∼1 dB in PSNR and ∼0.01 in SSIM under unseen scenarios. Our code is available at https://github.com/sriprabhar/SHFormer.

Objective classification biomarkers that are developed using resting-state functional magnetic resonance imaging (rs-fMRI) data are expected to contribute to more effective treatment for psychiatric disorders. Unfortunately, no widely accepted biomarkers are available at present, partially because of the large variety of analysis pipelines for their development. In this study, we comprehensively evaluated analysis pipelines using a large-scale, multi-site fMRI dataset for major depressive disorder (MDD). We explored combinations of options in four sub-processes of the analysis pipelines: six types of brain parcellation, four types of functional connectivity (FC) estimations, three types of site-difference harmonization, and five types of machine-learning methods. A total of 360 different MDD classification biomarkers were constructed using the SRPBS dataset acquired with unified protocols (713 participants from four sites) as the discovery dataset, and datasets from other projects acquired with heterogeneous protocols (449 participants from four sites) were used for independent validation. We repeated the procedure after swapping the roles of the two datasets to identify superior pipelines, regardless of the discovery dataset. The classification results of the top 10 biomarkers showed high similarity, and weight similarity was observed between eight of the biomarkers, except for two that used both data-driven parcellation and FC computation. We applied the top 10 pipelines to the datasets of other psychiatric disorders (autism spectrum disorder and schizophrenia), and eight of the biomarkers exhibited sufficient classification performance for both disorders. Our results will be useful for establishing a standardized pipeline for classification biomarkers.

Myocardial perfusion imaging (MPI) using single-photon emission computed tomography (SPECT) is a well-established modality for noninvasive diagnostic assessment of coronary artery disease (CAD). However, the time-consuming and experience-dependent visual interpretation of SPECT images remains a limitation in the clinic. We aimed to develop advanced models to diagnose CAD using different supervised and semi-supervised deep learning (DL) algorithms and training strategies, including transfer learning and data augmentation, with SPECT-MPI and invasive coronary angiography (ICA) as standard of reference. A total of 940 patients who underwent SPECT-MPI were enrolled (281 patients included ICA). Quantitative perfusion SPECT (QPS) was used to extract polar maps of rest and stress states. We defined two different tasks, including (1) Automated CAD diagnosis with expert reader (ER) assessment of SPECT-MPI as reference, and (2) CAD diagnosis from SPECT-MPI based on reference ICA reports. In task 2, we used 6 strategies for training DL models. We implemented 13 different DL models along with 4 input types with and without data augmentation (WAug and WoAug) to train, validate, and test the DL models (728 models). One hundred patients with ICA as standard of reference (the same patients in task 1) were used to evaluate models per vessel and per patient. Metrics, such as the area under the receiver operating characteristics curve (AUC), accuracy, sensitivity, specificity, precision, and balanced accuracy were reported. DeLong and pairwise Wilcoxon rank sum tests were respectively used to compare models and strategies after 1000 bootstraps on the test data for all models. We also compared the performance of our best DL model to ER's diagnosis. In task 1, DenseNet201 Late Fusion (AUC = 0.89) and ResNet152V2 Late Fusion (AUC = 0.83) models outperformed other models in per-vessel and per-patient analyses, respectively. In task 2, the best models for CAD prediction based on ICA were Strategy 3 (a combination of ER- and ICA-based diagnosis in train data), WoAug InceptionResNetV2 EarlyFusion (AUC = 0.71), and Strategy 5 (semi-supervised approach) WoAug ResNet152V2 EarlyFusion (AUC = 0.77) in per-vessel and per-patient analyses, respectively. Moreover, saliency maps showed that models could be helpful for focusing on relevant spots for decision making. Our study confirmed the potential of DL-based analysis of SPECT-MPI polar maps in CAD diagnosis. In the automation of ER-based diagnosis, models' performance was promising showing accuracy close to expert-level analysis. It demonstrated that using different strategies of data combination, such as including those with and without ICA, along with different training methods, like semi-supervised learning, can increase the performance of DL models. The proposed DL models could be coupled with computer-aided diagnosis systems and be used as an assistant to nuclear medicine physicians to improve their diagnosis and reporting, but only in the LAD territory. Not applicable.

Retrospective cohort study. To quantify the pathology of the spine in patients with scoliosis through one-dimensional feature analysis. Biplanar radiograph (EOS) imaging is a low-dose technology offering high-resolution spinal curvature measurement, crucial for assessing scoliosis severity and guiding treatment decisions. Machine learning (ML) algorithms, utilizing one-dimensional image features, can enable automated Cobb angle classification, improving accuracy and efficiency in scoliosis evaluation while reducing the need for manual measurements, thus supporting clinical decision-making. This study used 816 annotated AP EOS spinal images with a spine segmentation mask and a 10° polynomial to represent curvature. Engineered features included the first and second derivatives, Fourier transform, and curve energy, normalized for robustness. XGBoost selected the top 32 features. The models classified scoliosis into multiple groups based on curvature degree, measured through Cobb angle. To address the class imbalance, stratified sampling, undersampling, and oversampling techniques were used, with 10-fold stratified K-fold cross-validation for generalization. An automatic grid search was used for hyperparameter optimization, with K-fold cross-validation (K=3). The top-performing model was Random Forest, achieving an ROC AUC of 91.8%. An accuracy of 86.1%, precision of 86.0%, recall of 86.0%, and an F1 score of 85.1% were also achieved. Of the three techniques used to address class imbalance, stratified sampling produced the best out-of-sample results. SHAP values were generated for the top 20 features, including spine curve length and linear regression error, with the most predictive features ranked at the top, enhancing model explainability. Feature engineering with classical ML methods offers an effective approach for classifying scoliosis severity based on Cobb angle ranges. The high interpretability of features in representing spinal pathology, along with the ease of use of classical ML techniques, makes this an attractive solution for developing automated tools to manage complex spinal measurements.

This narrative review provides an overview of the use, differences, and clinical impact of current methods for kidney stone volume assessment. The different approaches to volume measurement are based on noncontrast computed tomography (NCCT). While volume measurement using formulas is sufficient for smaller stones, it tends to overestimate volume for larger or irregularly shaped calculi. In contrast, software-based segmentation significantly improves accuracy and reproducibility, and artificial intelligence based volumetry additionally shows excellent agreement with reference standards while reducing observer variability and measurement time. Moreover, specific CT preparation protocols may further enhance image quality and thus improve measurement accuracy. Clinically, stone volume has proven to be a superior predictor of stone-related events during follow-up, spontaneous stone passage under conservative management, and stone-free rates after shockwave lithotripsy (SWL) and ureteroscopy (URS) compared to linear measurements. Although manual measurement remains practical, its accuracy diminishes for complex or larger stones. Software-based segmentation and volumetry offer higher precision and efficiency but require established standards and broader access to dedicated software for routine clinical use.

Accurate prediction of adverse pathology (AP) in prostate cancer (PCa) patients is crucial for formulating effective treatment strategies. This study aims to develop and evaluate a multimodal deep learning model based on [¹⁸F]PSMA-1007 PET/CT and multiparametric MRI (mpMRI) to predict the presence of AP, and investigate whether the model that integrates [¹⁸F]PSMA-1007 PET/CT and mpMRI outperforms the individual PET/CT or mpMRI models in predicting AP. 341 PCa patients who underwent radical prostatectomy (RP) with mpMRI and PET/CT scans were retrospectively analyzed. We generated deep learning signature from mpMRI and PET/CT with a multimodal deep learning model (MPC) based on convolutional neural networks and transformer, which was subsequently incorporated with clinical characteristics to construct an integrated model (MPCC). These models were compared with clinical models and single mpMRI or PET/CT models. The MPCC model showed the best performance in predicting AP (AUC, 0.955 [95% CI: 0.932-0.975]), which is higher than MPC model (AUC, 0.930 [95% CI: 0.901-0.955]). The performance of the MPC model is better than that of single PET/CT (AUC, 0.813 [95% CI: 0.780-0.845]) or mpMRI (AUC, 0.865 [95% CI: 0.829-0.901]). Additionally, MPCC model is also effective in predicting single adverse pathological features. The deep learning model that integrates mpMRI and [¹⁸F]PSMA-1007 PET/CT enhances the predictive capabilities for the presence of AP in PCa patients. This improvement aids physicians in making informed preoperative decisions, ultimately enhancing patient prognosis.

To investigate the feasibility of characterizing tumor heterogeneity in breast cancer ultrasound images using habitat analysis technology and establish a radiomics machine learning model for predicting response to neoadjuvant chemotherapy (NAC). Ultrasound images from patients with pathologically confirmed breast cancer who underwent neoadjuvant therapy at our institution between July 2021 and December 2023 were retrospectively reviewed. Initially, the region of interest was delineated and segmented into multiple habitat areas using local feature delineation and cluster analysis techniques. Subsequently, radiomics features were extracted from each habitat area to construct 3 machine learning models. Finally, the model's efficacy was assessed through operating characteristic (ROC) curve analysis, decision curve analysis (DCA), and calibration curve evaluation. A total of 945 patients were enrolled, with 333 demonstrating a favorable response to NAC and 612 exhibiting an unfavorable response to NAC. Through the application of habitat analysis techniques, 3 distinct habitat regions within the tumor were identified. Subsequently, a predictive model was developed by incorporating 19 radiomics features, and all 3 machine learning models demonstrated excellent performance in predicting treatment outcomes. Notably, extreme gradient boosting (XGBoost) exhibited superior performance with an area under the curve (AUC) of 0.872 in the training cohort and 0.740 in the testing cohort. Additionally, DCA and calibration curves were employed for further evaluation. The habitat analysis technique effectively distinguishes distinct biological subregions of breast cancer, while the established radiomics machine learning model predicts NAC response by forecasting residual cancer burden (RCB) classification.

To develop CT-based machine learning radiomics models used for the diagnosis of dysthyroid optic neuropathy (DON). This is a retrospective study included 57 patients (114 orbits) diagnosed with thyroid-associated ophthalmopathy (TAO) at the Beijing Tongren Hospital between December 2019 and June 2023. CT scans, medical history, examination results, and clinical data of the participants were collected. DON was diagnosed based on clinical manifestations and examinations. The DON orbits and non-DON orbits were then divided into a training set and a test set at a ratio of approximately 7:3. The 3D slicer software was used to identify the volumes of interest (VOI). Radiomics features were extracted using the Pyradiomics and selected by t-test and least absolute shrinkage and selection operator (LASSO) regression algorithm with 10-fold cross-validation. Machine-learning models, including random forest (RF) model, support vector machine (SVM) model, and logistic regression (LR) model were built and validated by receiver operating characteristic (ROC) curves, area under the curves (AUC) and confusion matrix-related data. The net benefit of the models is shown by the decision curve analysis (DCA). We extracted 107 features from the imaging data, representing various image information of the optic nerve and surrounding orbital tissues. Using the LASSO method, we identified the five most informative features. The AUC ranged from 0.77 to 0.80 in the training set and the AUC of the RF, SVM and LR models based on the features were 0.86, 0.80 and 0.83 in the test set, respectively. The DeLong test showed there was no significant difference between the three models (RF model vs SVM model: <i>p</i> = .92; RF model vs LR model: <i>p</i> = .94; SVM model vs LR model: <i>p</i> = .98) and the models showed optimal clinical efficacy in DCA. The CT-based machine learning radiomics analysis exhibited excellent ability to diagnose DON and may enhance diagnostic convenience.