Publications by authors named "Akshay Chaudhari"
Recent advancements in medicine have confirmed that brain disorders often comprise multiple subtypes of mechanisms, developmental trajectories, or severity levels. Such heterogeneity is often associated with demographic aspects (e.g.
View Article and Find Full Text PDFArticle Synopsis
- This study investigates potential biases in both human readings of medical images and AI tools trained on human data, specifically focusing on knee osteoarthritis grading.
- Researchers used a dataset of 50 patients for external validation and a larger cohort of 8,273 to analyze the performance of an FDA-approved AI tool.
- Findings indicated that the AI tool displayed non-uniformity in disease grading, showing discrepancies of 20-22% and 13.6% in different patient datasets, but its overall accuracy was comparable to experienced radiologists without evidence of age or sex bias.
Article Synopsis
- Eccentric training via Nordic hamstring exercises (NHE) is effective in preventing hamstring strains by promoting changes in muscle structure, specifically increasing muscle fascicle length and adding sarcomeres in series within the muscle fibers.
- In a study with 12 participants, after 9 weeks of NHE training, the biceps femoris long-head (BFlh) showed significant improvements, including a 19% and 33% increase in fascicle length in the central and distal regions, respectively, along with a 40% increase in knee flexion strength.
- Following a 3-week period of no training (detraining), muscle adaptations such as fascicle length
Article Synopsis
- Analyzing the shapes of tissues and organs is crucial for diagnosing diseases like osteoarthritis, which affects many Americans; a new dataset called ShapeMed-Knee has been introduced to support this analysis.
- ShapeMed-Knee contains 9,376 high-resolution 3D shapes of femur bones and cartilage, along with benchmarks for accuracy and clinical prediction tasks, enhancing the understanding of osteoarthritis.
- The authors developed a cutting-edge hybrid neural shape model using ShapeMed-Knee that significantly improves reconstruction accuracy and accurately predicts localized osteoarthritis features, with plans to make the dataset, code, and benchmarks publicly available.
Article Synopsis
- AI models for medical imaging need large and diverse datasets, which are hard to obtain due to privacy concerns and data sharing issues.
- Synthetic medical imaging data, generated by AI, can help address these shortages while allowing for new applications and professional training.
- However, using synthetic data raises challenges related to realism, evaluation of model performance, high costs, and the need for updated regulations to ensure ethical use, highlighting the importance of collaboration between regulatory bodies, physicians, and AI developers.
Objective: A fully automated laminar cartilage composition (MRI-based T2) analysis method was technically and clinically validated by comparing radiographically normal knees with (CL-JSN) and without contra-lateral joint space narrowing or other signs of radiographic osteoarthritis (OA, CL-noROA).
Materials And Methods: 2D U-Nets were trained from manually segmented femorotibial cartilages (n = 72) from all 7 echoes (All), or from the 1st echo only (1) of multi-echo-spin-echo (MESE) MRIs acquired by the Osteoarthritis Initiative (OAI). Because of its greater accuracy, only the All U-Net was then applied to knees from the OAI healthy reference cohort (n = 10), CL-JSN (n = 39), and (1:1) matched CL-noROA knees (n = 39) that all had manual expert segmentation, and to 982 non-matched CL-noROA knees without expert segmentation.
Article Synopsis
- The lack of high-quality medical imaging datasets can be addressed using machine learning to create diverse images that accurately depict medical conditions and concepts.
- Current large vision-language models struggle because they're mainly trained on natural images, making their generated medical images less accurate.
- A new domain-adaptation method combines existing chest X-ray datasets and radiology reports to adapt a model, allowing it to produce synthetic medical images that are visually credible and can be tailored using specific medical text prompts.
Article Synopsis
- - The article outlines a proposed MRI acquisition protocol for clinical trials involving knee osteoarthritis, focusing on both early and advanced stages of the disease while supporting automated data analysis for specific imaging endpoints.
- - A comprehensive literature review and expert input were utilized to determine optimal MRI techniques, ensuring that the protocols can be executed within 30 minutes on standard clinical equipment.
- - The authors emphasize the importance of acquiring high-quality, longitudinal MRIs that include specific sequences for analyzing cartilage and synovitis, aiming to enhance scientific research in disease progression and treatment efficacy.
Article Synopsis
- Human pose estimation models often struggle with accuracy when detecting joint kinematics due to sparse keypoint detection, but OpenCap aims to improve this with a new deep learning model called the marker enhancer.
- A larger and more diverse training dataset, compiled from motion capture data involving 1,176 subjects, has been created to enhance the model's performance on various movements, even those not included in the training set.
- The updated marker enhancer has shown significant improvements in kinematic accuracy for benchmark and unseen movements compared to previous versions, making OpenCap a more reliable tool for researchers needing accurate movement measurements.
Deep learning (DL) has recently emerged as a pivotal technology for enhancing magnetic resonance imaging (MRI), a critical tool in diagnostic radiology. This review paper provides a comprehensive overview of recent advances in DL for MRI reconstruction, and focuses on various DL approaches and architectures designed to improve image quality, accelerate scans, and address data-related challenges. It explores end-to-end neural networks, pre-trained and generative models, and self-supervised methods, and highlights their contributions to overcoming traditional MRI limitations.
View Article and Find Full Text PDFArticle Synopsis
- - Over 85 million CT scans are done annually in the US, with a significant portion focused on the abdomen, highlighting a need for efficient interpretation methods due to a shortage of radiologists.
- - To address this, researchers introduced Merlin, a 3D Vision Language Model (VLM) that uses both electronic health records and radiology reports for training without the need for manual annotations, utilizing a vast clinical dataset of millions of images and codes.
- - Merlin was evaluated on various tasks, including chronic disease prediction and report generation, showing better performance than current methods, demonstrating its potential to support radiologists in their work.
Applications of Artificial Intelligence for Pediatric Cancer Imaging.
AJR Am J Roentgenol
August 2024
Artificial intelligence (AI) is transforming the medical imaging of adult patients. However, its utilization in pediatric oncology imaging remains constrained, in part due to the inherent scarcity of data associated with childhood cancers. Pediatric cancers are rare, and imaging technologies are evolving rapidly, leading to insufficient data of a particular type to effectively train these algorithms.
View Article and Find Full Text PDFArticle Synopsis
- Cryogenic electron tomography (cryoET) is an advanced imaging technique that captures detailed 3D images of biological specimens but struggles with data collection limitations like the missing wedge problem.
- Recent advancements using supervised deep learning methods, particularly convolutional neural networks (CNNs), have helped improve cryoET quality but require substantial pretraining, which can lead to inaccuracies when training data is limited.
- To address these issues, a new unsupervised learning approach using coordinate networks (CNs) has been proposed, significantly speeding up reconstruction times and enhancing image quality without needing pretraining, as demonstrated by improved shape completion and fewer artifacts in experimental results.
Article Synopsis
- - The text discusses the importance of analyzing the shapes of tissues and organs for diagnosing diseases, focusing on osteoarthritis, which affects a significant number of people in the U.S.
- - A new 3D shape dataset called ShapeMed-Knee has been introduced, containing 9,376 high-resolution models of the femur bone and cartilage, along with benchmarks for accuracy and clinical prediction tasks.
- - The study presents a hybrid neural shape model that outperforms existing models in accuracy and the ability to predict features related to osteoarthritis, aiming to improve medical diagnostics while providing open access to the dataset and tools.
Article Synopsis
- - Cryogenic electron tomography (cryoET) provides high-resolution 3D imaging of biological samples, but it struggles with the "missing wedge" problem which affects data quality due to limited collection angles.
- - Recent advancements in supervised deep learning, particularly convolutional neural networks (CNNs), have improved cryoET but often rely heavily on pretraining, which can lead to errors when training data is limited.
- - The proposed unsupervised learning method using coordinate networks (CNs) eliminates the need for pretraining, significantly speeds up reconstruction times, and improves image quality by reducing artifacts, offering insights on both supervised and unsupervised learning for better cryoET methods.
Medical artificial intelligence (AI) has tremendous potential to advance healthcare by supporting and contributing to the evidence-based practice of medicine, personalizing patient treatment, reducing costs, and improving both healthcare provider and patient experience. Unlocking this potential requires systematic, quantitative evaluation of the performance of medical AI models on large-scale, heterogeneous data capturing diverse patient populations. Here, to meet this need, we introduce MedPerf, an open platform for benchmarking AI models in the medical domain.
View Article and Find Full Text PDFArticle Synopsis
- - The study investigates the reproducibility of knee cartilage T mapping using a fast technique called qDESS, focusing on its ability to identify joints at risk for osteoarthritis.
- - Researchers evaluated two methods for analyzing cartilage: manual segmentation of specific regions and automatic segmentation through a deep-learning tool, assessing test-retest performance over different time intervals.
- - The analysis revealed that all cartilage regions demonstrated good reproducibility, allowing for better profiling of this biomarker's technical performance in clinical assessments.
Article Synopsis
- Researchers developed and validated an open-source AI algorithm to detect different contrast phases in abdominal CT scans, using data from 739 exams across 200 patients.
- The algorithm achieved high accuracy rates of 92.3% for internal testing and 90.1% for external validation, indicating strong performance in identifying non-contrast, arterial, venous, and delayed phases.
- The study confirms the algorithm's effectiveness and potential for clinical applications, enhancing how medical professionals interpret CT scans for improved patient outcomes.
Article Synopsis
- A study was conducted utilizing deep learning techniques to analyze abdominal CT scans for skeletal muscle metrics and their relationships with various medical conditions in a large North American cohort.
- The analysis included 17,646 adults and found significant associations between skeletal muscle index (SMI) and skeletal muscle density (SMD) with numerous medical phenotypes, including both previously known and unreported connections.
- Key findings showed that higher SMI correlated with a decrease in certain conditions like cardiac dysrhythmias and epilepsy, while higher SMD was associated with lower rates of decubitus ulcers and sleep disorders, emphasizing the potential of CT-derived muscle metrics in medical assessments.
Article Synopsis
- Analyzing large amounts of textual data from electronic health records can overwhelm clinicians, affecting their time management.
- This study tested eight large language models (LLMs) on various clinical summarization tasks, finding that their adapted versions performed comparably or better than expert medical summaries in many cases.
- The research indicates that integrating LLMs into clinical processes might reduce documentation workload, helping doctors dedicate more time to patient care.
Article Synopsis
- The significance of imaging biomarkers has grown in recent years, particularly in their clinical and research applications related to body composition.
- Key imaging techniques such as dual-energy X-ray absorptiometry, computed tomography, magnetic resonance imaging, and ultrasonography are used to assess bone, muscle, and fat tissues.
- Understanding the specific terminology related to clinical and imaging practices is essential for effectively utilizing these biomarkers in health assessments.
Article Synopsis
- Recent deep learning techniques can enhance kinetic assessments using IMU data, but they usually need a lot of labeled ground reaction force (GRF) data, which is often scarce.
- The researchers propose using self-supervised learning (SSL) to pre-train deep learning models with large IMU datasets, improving GRF estimation accuracy and making better use of available data.
- The study shows that SSL pre-training can significantly boost GRF estimation accuracy with minimal labeled data, suggesting that it can make IMU-driven assessments more practical and accessible.*
Article Synopsis
- Recent advancements in deep learning techniques can enhance IMU-driven kinetic assessment but require substantial amounts of ground reaction force (GRF) data for training; self-supervised learning (SSL) can help utilize large IMU datasets for pre-training models, improving accuracy and data efficiency in GRF estimation.* -
- The study involved masking parts of IMU data and training a transformer model to predict the masked sections, comparing various masking ratios across different datasets (real, synthetic, and combined) to optimize performance.* -
- Results showed that SSL pre-training significantly increased the accuracy of GRF estimation during walking tasks, allowing models to achieve similar accuracy with much less labeled data, and suggesting an optimal masking ratio of 6.25-
Article Synopsis
- The study aimed to analyze how age affects injury and adaptation patterns in the elbows of overhead throwing athletes specifically related to valgus extension overload (VEO).
- A total of 86 throwing athletes were categorized into three age groups (≤16, 17-19, and ≥20 years) and compared to 23 non-athlete controls using MR imaging to assess ulnar collateral ligament (UCL) thickness and subchondral sclerosis.
- Findings revealed that younger athletes had more apophyseal and stress injuries, while older athletes experienced more soft tissue injuries, with significant anatomical differences observed in UCL thickness and sclerosis between athletes and controls, as well as among different age groups of athletes.