Privacy-preserving large language models (PP-LLMs) hold potential for assisting clinicians with documentation. We evaluated a PP-LLM to improve the clinical information on radiology request forms for musculoskeletal magnetic resonance imaging (MRI) and to automate protocoling, which ensures that the most appropriate imaging is performed. The present retrospective study included musculoskeletal MRI radiology request forms that had been randomly collected from June to December 2023. Studies without electronic medical record (EMR) entries were excluded. An institutional PP-LLM (Claude Sonnet 3.5) augmented the original radiology request forms by mining EMRs, and, in combination with rule-based processing of the LLM outputs, suggested appropriate protocols using institutional guidelines. Clinical information on the original and PP-LLM radiology request forms were compared with use of the RI-RADS (Reason for exam Imaging Reporting and Data System) grading by 2 musculoskeletal (MSK) radiologists independently (MSK1, with 13 years of experience, and MSK2, with 11 years of experience). These radiologists established a consensus reference standard for protocoling, against which the PP-LLM and of 2 second-year board-certified radiologists (RAD1 and RAD2) were compared. Inter-rater reliability was assessed with use of the Gwet AC1, and the percentage agreement with the reference standard was calculated. Overall, 500 musculoskeletal MRI radiology request forms were analyzed for 407 patients (202 women and 205 men with a mean age [and standard deviation] of 50.3 ± 19.5 years) across a range of anatomical regions, including the spine/pelvis (143 MRI scans; 28.6%), upper extremity (169 scans; 33.8%) and lower extremity (188 scans; 37.6%). Two hundred and twenty-two (44.4%) of the 500 MRI scans required contrast. The clinical information provided in the PP-LLM-augmented radiology request forms was rated as superior to that in the original requests. Only 0.4% to 0.6% of PP-LLM radiology request forms were rated as limited/deficient, compared with 12.4% to 22.6% of the original requests (p < 0.001). Almost-perfect inter-rater reliability was observed for LLM-enhanced requests (AC1 = 0.99; 95% confidence interval [CI], 0.99 to 1.0), compared with substantial agreement for the original forms (AC1 = 0.62; 95% CI, 0.56 to 0.67). For protocoling, MSK1 and MSK2 showed almost-perfect agreement on the region/coverage (AC1 = 0.96; 95% CI, 0.95 to 0.98) and contrast requirement (AC1 = 0.98; 95% CI, 0.97 to 0.99). Compared with the consensus reference standard, protocoling accuracy for the PP-LLM was 95.8% (95% CI, 94.0% to 97.6%), which was significantly higher than that for both RAD1 (88.6%; 95% CI, 85.8% to 91.4%) and RAD2 (88.2%; 95% CI, 85.4% to 91.0%) (p < 0.001 for both). Musculoskeletal MRI request form augmentation with an institutional LLM provided superior clinical information and improved protocoling accuracy compared with clinician requests and non-MSK-trained radiologists. Institutional adoption of such LLMs could enhance the appropriateness of MRI utilization and patient care. Diagnostic Level III. See Instructions for Authors for a complete description of levels of evidence.

Biomedical image classification is of paramount importance in enhancing diagnostic precision and improving patient outcomes across diverse medical disciplines. In recent years, the advent of deep learning methodologies has significantly transformed this domain by facilitating notable advancements in image analysis and classification endeavors. This paper provides a thorough overview of the application of deep learning techniques in biomedical image classification, encompassing various types of healthcare data, including medical images derived from modalities such as mammography, histopathology, and radiology. A detailed discourse on deep learning architectures, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and advanced models such as generative adversarial networks (GANs), is presented. Additionally, we delineate the distinctions between supervised, unsupervised, and reinforcement learning approaches, along with their respective roles within the context of biomedical imaging. This study systematically investigates 50 deep learning methodologies employed in the healthcare sector, elucidating their effectiveness in various tasks, including disease detection, image segmentation, and classification. It particularly emphasizes models that have been trained on publicly available datasets, thereby highlighting the significant role of open-access data in fostering advancements in AI-driven healthcare innovations. Furthermore, this review accentuates the transformative potential of deep learning in the realm of biomedical image analysis and delineates potential avenues for future research within this rapidly evolving field.

Hydatid cysts, caused by Echinococcus granulosus, form progressively enlarging fluid-filled cysts in organs like the liver and lungs, posing significant public health risks through severe complications or death. This study presents a novel deep feature generation framework utilizing vision transformer models (ViT-DFG) to enhance the classification accuracy of hydatid cyst types. The proposed framework consists of four phases: image preprocessing, feature extraction using vision transformer models, feature selection through iterative neighborhood component analysis, and classification, where the performance of the ViT-DFG model was evaluated and compared across different classifiers such as k-nearest neighbor and multi-layer perceptron (MLP). Both methods were evaluated independently to assess classification performance from different approaches. The dataset, comprising five cyst types, was analyzed for both five-class and three-class classification by grouping the cyst types into active, transition, and inactive categories. Experimental results showed that the proposed VIT-DFG method achieves higher accuracy than existing methods. Specifically, the ViT-DFG framework attained an overall classification accuracy of 98.10% for the three-class and 95.12% for the five-class classifications using 5-fold cross-validation. Statistical analysis through one-way analysis of variance (ANOVA), conducted to evaluate significant differences between models, confirmed significant differences between the proposed framework and individual vision transformer models ( <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>p</mi> <mo><</mo> <mn>0.05</mn></mrow> </math> ). These results highlight the effectiveness of combining multiple vision transformer architectures with advanced feature selection techniques in improving classification performance. The findings underscore the ViT-DFG framework's potential to advance medical image analysis, particularly in hydatid cyst classification, while offering clinical promise through automated diagnostics and improved decision-making.

Education for medical imaging technologists or radiographers in regional and rural areas often faces significant challenges due to limited financial, technological, and teaching resources. Generative AI presents a promising solution to overcome these barriers and support the professional development of radiographers. This pilot study aimed to evaluate the educational value of an in-house AI-based imaging simulation tool designed to generate clinically relevant medical images for professional training purposes. In July 2023, a professional development lecture featuring AI-generated clinical imaging content was delivered to students (N = 122/130) and recent graduates (N = 155/532), alongside a pre-lecture survey. Following the session, participants completed a questionnaire comprising structured and open-ended items to assess their understanding, perceptions, and interest in AI within medical imaging education. Survey results indicated that both students and graduates possessed a foundational awareness of AI applications in medical imaging. Graduates demonstrated significantly higher expectations for clinical realism in AI-generated simulations, likely reflecting their clinical experience. Although the simulator's current capabilities are limited in replicating complex diagnostic imaging, participants acknowledged its pedagogical value, particularly in supporting basic anatomical education. Approximately 50% of respondents expressed interest in further developing their AI knowledge and contributing to the research and development of AI-based educational tools. AI-driven imaging simulation tools have the potential to enhance radiography education and reduce teaching barriers. While further development is needed to improve clinical fidelity, such tools can play a valuable role in foundational training and foster learner engagement in AI innovation.

Reduced muscle mass and function are associated with increased morbidity, and mortality. Ultrasound, despite being cost-effective and portable, is still underutilized in muscle trophism assessment due to its reliance on operator expertise and measurement variability. This proof-of-concept study aimed to overcome these limitations by developing a deep learning model that predicts muscle density, as assessed by CT, using Ultrasound data, exploring the feasibility of a novel Ultrasound-based parameter for muscle trophism.A sample of adult participants undergoing CT examination in our institution's emergency department between May 2022 and March 2023 was enrolled in this single-center study. Ultrasound examinations were performed with a L11-3 MHz probe. The rectus abdominis muscles, selected as target muscles, were scanned in the transverse plane, recording an Ultrasound image per side. For each participant, the same operator calculated the average target muscle density in Hounsfield Units from an axial CT slice closely matching the Ultrasound scanning plane.The final dataset included 1090 Ultrasound images from 551 participants (mean age 67 ± 17, 323 males). A deep learning model was developed to classify Ultrasound images into three muscle-density classes based on CT values. The model achieved promising performance, with a categorical accuracy of 70% and AUC values of 0.89, 0.79, and 0.90 across the three classes.This observational study introduces an innovative approach to automated muscle trophism assessment using Ultrasound imaging. Future efforts should focus on external validation in diverse populations and clinical settings, as well as expanding its application to other muscles.

Deep learning has revolutionized medical imaging, improving tasks like image segmentation, detection, and classification, often surpassing human accuracy. However, the training of effective diagnostic models is hindered by two major challenges: the need for large datasets for each task and privacy laws restricting the sharing of medical data. Multi-task learning (MTL) addresses the first challenge by enabling a single model to perform multiple tasks, though convolution-based MTL models struggle with contextualizing global features. Federated learning (FL) helps overcome the second challenge by allowing models to train collaboratively without sharing data, but traditional methods struggle to aggregate stable feature maps due to the permutation-invariant nature of neural networks. To tackle these issues, we propose MTMedFormer, a transformer-based multi-task medical imaging model. We leverage the transformers' ability to learn task-agnostic features using a shared encoder and utilize task-specific decoders for robust feature extraction. By combining MTL with a hybrid loss function, MTMedFormer learns distinct diagnostic tasks in a synergistic manner. Additionally, we introduce a novel Bayesian federation method for aggregating multi-task imaging models. Our results show that MTMedFormer outperforms traditional single-task and MTL models on mammogram and pneumonia datasets, while our Bayesian federation method surpasses traditional methods in image segmentation.

The histological grade of hepatocellular carcinoma (HCC) is an important factor associated with early tumor recurrence and prognosis after surgery. Developing a valuable tool to assess this grade is essential for treatment. This study aimed to evaluate the feasibility and efficacy of a deep learning-based three-dimensional super-resolution (SR) magnetic resonance imaging radiomics model for predicting the pathological grade of HCC. A total of 197 HCC patients were included and divided into a training cohort (n = 157) and a testing cohort (n = 40). Three-dimensional SR technology based on deep learning was used to obtain SR hepatobiliary phase (HBP) images from normal-resolution (NR) HBP images. High-dimensional quantitative features were extracted from manually segmented volumes of interest in NRHBP and SRHBP images. The gradient boosting, light gradient boosting machine, and support vector machine were used to develop three-class (well-differentiated vs. moderately differentiated vs. poorly differentiated) and binary radiomics (well-differentiated vs. moderately and poorly differentiated) models, and the predictive performance of these models was evaluated using several measures. All the three-class models using SRHBP images had higher area under the curve (AUC) values than those using NRHBP images. The binary classification models developed with SRHBP images also outperformed those with NRHBP images in distinguishing moderately and poorly differentiated HCC from well-differentiated HCC (AUC = 0.849, sensitivity = 77.8%, specificity = 76.9%, accuracy = 77.5% vs. AUC = 0.603, sensitivity = 48.1%, specificity = 76.9%, accuracy = 57.5%; p = 0.039). Decision curve analysis revealed the clinical value of the models. Deep learning-based three-dimensional SR technology may improve the performance of radiomics models using HBP images for predicting the preoperative pathological grade of HCC.

Determining age from dental remains is vital in forensic investigations, aiding in victim identification and anthropological research. Our framework uses a two-step pipeline: tooth detection followed by age estimation, based on either canine tooth images alone or combined with sex information. The dataset included 2,587 radiographs from 1,004 patients (691 females, 313 males) aged 13.42-85.45 years. The YOLOv8-Nano model achieved exceptional performance in detecting canine teeth, with an F1 score of 0.994, a 98.94% detection success rate, and accurate numbering of all detected teeth. For age estimation, we implemented four convolutional neural network architectures: ResNet-18, DenseNet-121, EfficientNet-B0, and MobileNetV3. Each model was trained to estimate age based on one of the four individual canine teeth (13, 23, 33, and 43). The models achieved median absolute errors ranging from 3.55 to 5.18 years. Incorporating sex as an additional input feature did not improve performance. Moreover, no significant differences in predictive accuracy were observed among the individual teeth. In conclusion, the proposed framework demonstrates potential as a robust and practical tool for forensic age estimation across diverse forensic contexts.

Medical imaging has become an essential tool in the diagnosis and treatment of various diseases, and provides critical insights through ultrasound, MRI, and X-ray modalities. Despite its importance, challenges remain in the accurate segmentation and classification of complex structures owing to factors such as low contrast, noise, and irregular anatomical shapes. This study addresses these challenges by proposing a novel hybrid deep learning model that integrates the strengths of Convolutional Autoencoders (CAE), UNet, and SegNet architectures. In the preprocessing phase, a Convolutional Autoencoder is used to effectively reduce noise while preserving essential image details, ensuring that the images used for segmentation and classification are of high quality. The ability of CAE to denoise images while retaining critical features enhances the accuracy of the subsequent analysis. The developed model employs UNet for multiscale feature extraction and SegNet for precise boundary reconstruction, with Dynamic Feature Fusion integrated at each skip connection to dynamically weight and combine the feature maps from the encoder and decoder. This ensures that both global and local features are effectively captured, while emphasizing the critical regions for segmentation. To further enhance the model's performance, the Hybrid Emperor Penguin Optimizer (HEPO) was employed for feature selection, while the Hybrid Vision Transformer with Convolutional Embedding (HyViT-CE) was used for the classification task. This hybrid approach allows the model to maintain high accuracy across different medical imaging tasks. The model was evaluated using three major datasets: brain tumor MRI, breast ultrasound, and chest X-rays. The results demonstrate exceptional performance, achieving an accuracy of 99.92% for brain tumor segmentation, 99.67% for breast cancer detection, and 99.93% for chest X-ray classification. These outcomes highlight the ability of the model to deliver reliable and accurate diagnostics across various medical contexts, underscoring its potential as a valuable tool in clinical settings. The findings of this study will contribute to advancing deep learning applications in medical imaging, addressing existing research gaps, and offering a robust solution for improved patient care.

Assessing tumor response to systemic therapies is one of the main applications of PET/CT. Routinely, only a small subset of index lesions out of multiple lesions is analyzed. However, this operator dependent selection may bias the results due to possible significant inter-metastatic heterogeneity of response to therapy. Automated, AI-based approaches for lesion tracking hold promise in enabling the analysis of many more lesions and thus providing a better assessment of tumor response. This work introduces a Siamese CNN approach for lesion tracking between PET/CT scans. Our approach is applied on the laborious task of tracking a high number of bone lesions in full-body baseline and follow-up [⁶⁸Ga]Ga- or [¹⁸F]F-PSMA PET/CT scans after two cycles of [¹⁷⁷Lu]Lu-PSMA therapy of metastatic castration resistant prostate cancer patients. Data preparation includes lesion segmentation and affine registration. Our algorithm extracts suitable lesion patches and forwards them into a Siamese CNN trained to classify the lesion patch pairs as corresponding or non-corresponding lesions. Experiments have been performed with different input patch types and a Siamese network in 2D and 3D. The CNN model successfully learned to classify lesion assignments, reaching an accuracy of 83 % in its best configuration with an AUC = 0.91. For corresponding lesions the pipeline accomplished lesion tracking accuracy of even 89 %. We proved that a CNN may facilitate the tracking of multiple lesions in PSMA PET/CT scans. Future clinical studies are necessary if this improves the prediction of the outcome of therapies.