Shreeram Athreya
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Shreeram Athreya has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- William Speier
University of California, Los Angeles
6 shared publications - Corey W. Arnold
University of California Los Angeles USA
5 shared publications - Vivek R. Sant
Section of Endocrine Surgery, Department of Surgery, UT Southwestern Medical Center, Dallas
4 shared publications - Ashwath Radhachandran
Department of Bioengineering, University of California Los Angeles, Los Angeles, CA, United States
2 shared publications - Kambiz Nael
Department of Radiology & Biomedical Imaging, University of California, San Francisco, California, USA
2 shared publications - Pradyumna Chari
Massachusetts Institute of Technology, Cambridge, MA, USA
2 shared publications
Latest Publications
- 25 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. PP (02780062) , 1
DSA-NRP: No-Reflow Prediction from Angiographic Perfusion Dynamics in Stroke EVT
Following successful large-vessel recanalization via endovascular thrombectomy (EVT) for acute ischemic stroke (AIS), some patients experience a complication known as no-reflow, defined by persistent microvascular hypoperfusion that undermines tissue recovery and worsens clinical outcomes. Although prompt identification is crucial, standard clinical practice relies on perfusion magnetic resonance imaging (MRI) within 24 hours post-procedure, delaying intervention. In this work, we introduce the first-ever machine learning (ML) framework to predict no-reflow immediately after EVT by leveraging previously unexplored intra-procedural digital subtraction angiography (DSA) sequences and clinical variables. Our retrospective analysis included AIS patients treated at UCLA Medical Center (2011–2024) who achieved favorable mTICI scores (2c or 3) and underwent pre- and post-procedure MRI. No-reflow was defined as a > 15% reduction in relative cerebral blood volume or flow within the infarct core compared to the contralateral hemisphere. From DSA sequences (anteroposterior and lateral views), we extracted statistical and temporal perfusion features from the target downstream territory to train ML classifiers for predicting no-reflow. Our preliminary results demonstrate that this novel method out-performed a clinical-features baseline (AUROC: 0.9330 vs. 0.7768 (p = 0.006)), suggesting that real-time DSA perfusion dynamics may encode clinically relevant information related to microvascular integrity. This approach establishes a preliminary foundation for immediate, accurate no-reflow prediction, enabling clinicians to proactively manage high-risk patients without reliance on delayed imaging, though it warrants validation in larger, independent cohorts.
- 25 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. PP (02780062) , 1
DSA-NRP: No-Reflow Prediction from Angiographic Perfusion Dynamics in Stroke EVT
Following successful large-vessel recanalization via endovascular thrombectomy (EVT) for acute ischemic stroke (AIS), some patients experience a complication known as no-reflow, defined by persistent microvascular hypoperfusion that undermines tissue recovery and worsens clinical outcomes. Although prompt identification is crucial, standard clinical practice relies on perfusion magnetic resonance imaging (MRI) within 24 hours post-procedure, delaying intervention. In this work, we introduce the first-ever machine learning (ML) framework to predict no-reflow immediately after EVT by leveraging previously unexplored intra-procedural digital subtraction angiography (DSA) sequences and clinical variables. Our retrospective analysis included AIS patients treated at UCLA Medical Center (2011–2024) who achieved favorable mTICI scores (2c or 3) and underwent pre- and post-procedure MRI. No-reflow was defined as a > 15% reduction in relative cerebral blood volume or flow within the infarct core compared to the contralateral hemisphere. From DSA sequences (anteroposterior and lateral views), we extracted statistical and temporal perfusion features from the target downstream territory to train ML classifiers for predicting no-reflow. Our preliminary results demonstrate that this novel method out-performed a clinical-features baseline (AUROC: 0.9330 vs. 0.7768 (p = 0.006)), suggesting that real-time DSA perfusion dynamics may encode clinically relevant information related to microvascular integrity. This approach establishes a preliminary foundation for immediate, accurate no-reflow prediction, enabling clinicians to proactively manage high-risk patients without reliance on delayed imaging, though it warrants validation in larger, independent cohorts.
- Open Access
Exploring Temporal Dynamics in No-Reflow Assessment
Background: For acute ischemic stroke (AIS) patients, successful endovascular thrombectomy (EVT) may not recover functional independence. One potential mechanism driving this disparity is no-reflow, a lack of perfusion despite recanalization of the large vessel occlusion (LVO). None of the published methods for assessing no-reflow have gained acceptance, and debate persists about isolating no-reflow from upstream hemodynamic anomalies as well as the temporal dynamics no-reflow assessment. This work conducts an exploratory analysis of the effect of time since EVT on the performance of different methods for assessing no-reflow. Methods: We identified a cohort of UCLA Health patients with successful LVO recanalization (≥ mTICI 2b) of a thrombus located in M1 or ICA. Patients were stratified into ‘early’ and ‘follow-up’ cohorts of < 24 and < 48 hours since EVT. A set of published methods using perfusion imaging for no-reflow detection were adapted to run semi-automatically for both cohorts. Results: We compiled 83 imaging studies from 64 patients. No-reflow rates varied from 4 . 2 % to 51 . 4 % by method, time since EVT, and mTICI. Follow-up no-reflow rates more closely match the original literature across all methods. mTICI 2b patients likely inflated no-reflow, as rates dropped approximately 50 % for mTICI 2c/3 cases. Early cohort inter-method agreement shows kappa scores as low as − 0 . 017 , but improved up to 0 . 72 in the follow-up cohort. Conclusion: No-reflow exhibits substantial dependence on time across all no-reflow detection methods studied. Inter-method agreement is good for methods using the same perfusion parameters, but each method is subject to confounding influences, motivating an aggregate method for robust evaluation. The temporal dynamics of no-reflow detection raises questions about imaging timing post-EVT for detection within the intervention window.
Open AccessExploring Temporal Dynamics in No-Reflow Assessment
Background: For acute ischemic stroke (AIS) patients, successful endovascular thrombectomy (EVT) may not recover functional independence. One potential mechanism driving this disparity is no-reflow, a lack of perfusion despite recanalization of the large vessel occlusion (LVO). None of the published methods for assessing no-reflow have gained acceptance, and debate persists about isolating no-reflow from upstream hemodynamic anomalies as well as the temporal dynamics no-reflow assessment. This work conducts an exploratory analysis of the effect of time since EVT on the performance of different methods for assessing no-reflow. Methods: We identified a cohort of UCLA Health patients with successful LVO recanalization (≥ mTICI 2b) of a thrombus located in M1 or ICA. Patients were stratified into ‘early’ and ‘follow-up’ cohorts of < 24 and < 48 hours since EVT. A set of published methods using perfusion imaging for no-reflow detection were adapted to run semi-automatically for both cohorts. Results: We compiled 83 imaging studies from 64 patients. No-reflow rates varied from 4 . 2 % to 51 . 4 % by method, time since EVT, and mTICI. Follow-up no-reflow rates more closely match the original literature across all methods. mTICI 2b patients likely inflated no-reflow, as rates dropped approximately 50 % for mTICI 2c/3 cases. Early cohort inter-method agreement shows kappa scores as low as − 0 . 017 , but improved up to 0 . 72 in the follow-up cohort. Conclusion: No-reflow exhibits substantial dependence on time across all no-reflow detection methods studied. Inter-method agreement is good for methods using the same perfusion parameters, but each method is subject to confounding influences, motivating an aggregate method for robust evaluation. The temporal dynamics of no-reflow detection raises questions about imaging timing post-EVT for detection within the intervention window.
- 1 May 2025
- Published by SAGE Publications in Thyroid®
- Vol. 35 (5) , 590-594
Combining Ultrasound Imaging and Molecular Testing in a Multimodal Deep Learning Model for Risk Stratification of Indeterminate Thyroid Nodules
Indeterminate cytology (Bethesda III and IV) represents 15–30% of biopsied thyroid nodules and require additional diagnostic testing. Molecular testing (MT) is a commonly used diagnostic tool that evaluatesmalignancy risk through next generation sequencing of fine needle aspiration (FNA) samples. While MT achieves high sensitivity (97–100%) in ruling out malignancy, its specificity and positive predictive value (PPV) remain relatively low. This study proposes a multimodal deep learning model that integrates ultrasound (US) imaging with MT to improve risk stratification by enhancing PPV while maintaining high sensitivity. Combining these modalities leverages complementary information from both molecular and imaging data, addressing limitations in current approaches and offering a robust framework for evaluating indeterminate nodules. We retrospectively analyzed 333 patients with indeterminate thyroid nodules (259 benign, 74 malignant) at UCLA Medical Center between 2016 and 2022. We evaluated four configurations: whole frame US images, 256 × 256 patches, 128 × 128 patches, and an ensemble model combining the first three configurations. The clinical baseline consisted of Bethesda cytology and MT results. Models were assessed using five fold cross validation stratified by surgical outcomes. The clinical baseline (Bethesda + MT) achieved an AUROC of 0.728 [0.68, 0.78] with sensitivity of 0.946 [0.88, 1.00], specificity of 0.664 [0.60, 0.73], and PPV of 0.448 [0.41, 0.48]. The proposed ensemble model demonstrated improved performance, achieving an AUROC of 0.831 [0.77, 0.89] with a sensitivity of 0.946 [0.88, 1.00], specificity of 0.703 [0.66, 0.75], and PPV of 0.477 [0.46, 0.50]. These improvements were statistically significant (p = 0.0008). Our multimodal model enhances MT performance by providing statistically significant improvements in PPV and specificity while maintaining high sensitivity. Our framework could be leveraged to reduce the number of benign thyroid resections in patients with indeterminate nodules. However, this study is limited by its single center dataset, lack of external validation, and the use of binarized MT outputs rather than granular malignancy risk probabilities. Future work should validate these findings across diverse populations and larger external datasets for more comprehensive risk stratification.
- 1 May 2025
- Published by SAGE Publications in Thyroid®
- Vol. 35 (5) , 590-594
Combining Ultrasound Imaging and Molecular Testing in a Multimodal Deep Learning Model for Risk Stratification of Indeterminate Thyroid Nodules
Indeterminate cytology (Bethesda III and IV) represents 15–30% of biopsied thyroid nodules and require additional diagnostic testing. Molecular testing (MT) is a commonly used diagnostic tool that evaluatesmalignancy risk through next generation sequencing of fine needle aspiration (FNA) samples. While MT achieves high sensitivity (97–100%) in ruling out malignancy, its specificity and positive predictive value (PPV) remain relatively low. This study proposes a multimodal deep learning model that integrates ultrasound (US) imaging with MT to improve risk stratification by enhancing PPV while maintaining high sensitivity. Combining these modalities leverages complementary information from both molecular and imaging data, addressing limitations in current approaches and offering a robust framework for evaluating indeterminate nodules. We retrospectively analyzed 333 patients with indeterminate thyroid nodules (259 benign, 74 malignant) at UCLA Medical Center between 2016 and 2022. We evaluated four configurations: whole frame US images, 256 × 256 patches, 128 × 128 patches, and an ensemble model combining the first three configurations. The clinical baseline consisted of Bethesda cytology and MT results. Models were assessed using five fold cross validation stratified by surgical outcomes. The clinical baseline (Bethesda + MT) achieved an AUROC of 0.728 [0.68, 0.78] with sensitivity of 0.946 [0.88, 1.00], specificity of 0.664 [0.60, 0.73], and PPV of 0.448 [0.41, 0.48]. The proposed ensemble model demonstrated improved performance, achieving an AUROC of 0.831 [0.77, 0.89] with a sensitivity of 0.946 [0.88, 1.00], specificity of 0.703 [0.66, 0.75], and PPV of 0.477 [0.46, 0.50]. These improvements were statistically significant (p = 0.0008). Our multimodal model enhances MT performance by providing statistically significant improvements in PPV and specificity while maintaining high sensitivity. Our framework could be leveraged to reduce the number of benign thyroid resections in patients with indeterminate nodules. However, this study is limited by its single center dataset, lack of external validation, and the use of binarized MT outputs rather than granular malignancy risk probabilities. Future work should validate these findings across diverse populations and larger external datasets for more comprehensive risk stratification.
- Open Access
CytoFM: The first cytology foundation model
Cytology is essential for cancer diagnostics and screening due to its minimally invasive nature. However, the development of robust deep learning models for digital cytology is challenging due to the heterogeneity in staining and preparation methods of samples, differences across organs, and the limited availability of large, diverse, annotated datasets. Developing a task-specific model for every cytology application is impractical and non-cytology-specific foundation models struggle to generalize to tasks in this domain where the emphasis is on cell morphology. To address these challenges, we introduce CytoFM, the first cytology self-supervised foundation model. Using iBOT, a self-supervised Vision Transformer (ViT) training framework incorporating masked image modeling and self-distillation, we pretrain CytoFM on a diverse collection of cytology datasets to learn robust, transferable representations. We evaluate CytoFM on multiple downstream cytology tasks, including breast cancer classification and cell type identification, using an attention-based multiple instance learning framework. Our results demonstrate that CytoFM performs better on two out of three downstream tasks than existing foundation models pretrained on histopathology (UNI) or natural images (iBOT-Imagenet). Visualizations of learned representations demonstrate our model is able to attend to cytologically relevant features. Despite a small pre-training dataset, CytoFM's promising results highlight the ability of task-agnostic pre-training approaches to learn robust and generalizable features from cytology data.
Open AccessCytoFM: The first cytology foundation model
Cytology is essential for cancer diagnostics and screening due to its minimally invasive nature. However, the development of robust deep learning models for digital cytology is challenging due to the heterogeneity in staining and preparation methods of samples, differences across organs, and the limited availability of large, diverse, annotated datasets. Developing a task-specific model for every cytology application is impractical and non-cytology-specific foundation models struggle to generalize to tasks in this domain where the emphasis is on cell morphology. To address these challenges, we introduce CytoFM, the first cytology self-supervised foundation model. Using iBOT, a self-supervised Vision Transformer (ViT) training framework incorporating masked image modeling and self-distillation, we pretrain CytoFM on a diverse collection of cytology datasets to learn robust, transferable representations. We evaluate CytoFM on multiple downstream cytology tasks, including breast cancer classification and cell type identification, using an attention-based multiple instance learning framework. Our results demonstrate that CytoFM performs better on two out of three downstream tasks than existing foundation models pretrained on histopathology (UNI) or natural images (iBOT-Imagenet). Visualizations of learned representations demonstrate our model is able to attend to cytologically relevant features. Despite a small pre-training dataset, CytoFM's promising results highlight the ability of task-agnostic pre-training approaches to learn robust and generalizable features from cytology data.
- Open Access
- 17 December 2024
- Published by JMIR Publications Inc. in JMIR Biomedical Engineering
- Vol. 9 (1) , e58911
Enhancing Ultrasound Image Quality Across Disease Domains: Application of Cycle-Consistent Generative Adversarial Network and Perceptual Loss
Background: Numerous studies have explored image processing techniques aimed at enhancing ultrasound images to narrow the performance gap between low-quality portable devices and high-end ultrasound equipment. These investigations often use registered image pairs created by modifying the same image through methods like down sampling or adding noise, rather than using separate images from different machines. Additionally, they rely on organ-specific features, limiting the models’ generalizability across various imaging conditions and devices. The challenge remains to develop a universal framework capable of improving image quality across different devices and conditions, independent of registration or specific organ characteristics. Objective: This study aims to develop a robust framework that enhances the quality of ultrasound images, particularly those captured with compact, portable devices, which are often constrained by low quality due to hardware limitations. The framework is designed to effectively process nonregistered ultrasound image pairs, a common challenge in medical imaging, across various clinical settings and device types. By addressing these challenges, the research seeks to provide a more generalized and adaptable solution that can be widely applied across diverse medical scenarios, improving the accessibility and quality of diagnostic imaging. Methods: A retrospective analysis was conducted by using a cycle-consistent generative adversarial network (CycleGAN) framework enhanced with perceptual loss to improve the quality of ultrasound images, focusing on nonregistered image pairs from various organ systems. The perceptual loss was integrated to preserve anatomical integrity by comparing deep features extracted from pretrained neural networks. The model’s performance was evaluated against corresponding high-resolution images, ensuring that the enhanced outputs closely mimic those from high-end ultrasound devices. The model was trained and validated using a publicly available, diverse dataset to ensure robustness and generalizability across different imaging scenarios. Results: The advanced CycleGAN framework, enhanced with perceptual loss, significantly outperformed the previous state-of-the-art, stable CycleGAN, in multiple evaluation metrics. Specifically, our method achieved a structural similarity index of 0.2889 versus 0.2502 (P<.001), a peak signal-to-noise ratio of 15.8935 versus 14.9430 (P<.001), and a learned perceptual image patch similarity score of 0.4490 versus 0.5005 (P<.001). These results demonstrate the model’s superior ability to enhance image quality while preserving critical anatomical details, thereby improving diagnostic usefulness. Conclusions: This study presents a significant advancement in ultrasound imaging by leveraging a CycleGAN model enhanced with perceptual loss to bridge the quality gap between images from different devices. By processing nonregistered image pairs, the model not only enhances visual quality but also ensures the preservation of essential anatomical structures, crucial for accurate diagnosis. This approach holds the potential to democratize high-quality ultrasound imaging, making it accessible through low-cost portable devices, thereby improving health care outcomes, particularly in resource-limited settings. Future research will focus on further validation and optimization for clinical use.
- 17 December 2024
- Published by JMIR Publications Inc. in JMIR Biomedical Engineering
- Vol. 9 (1) , e58911
Open AccessEnhancing Ultrasound Image Quality Across Disease Domains: Application of Cycle-Consistent Generative Adversarial Network and Perceptual Loss
Background: Numerous studies have explored image processing techniques aimed at enhancing ultrasound images to narrow the performance gap between low-quality portable devices and high-end ultrasound equipment. These investigations often use registered image pairs created by modifying the same image through methods like down sampling or adding noise, rather than using separate images from different machines. Additionally, they rely on organ-specific features, limiting the models’ generalizability across various imaging conditions and devices. The challenge remains to develop a universal framework capable of improving image quality across different devices and conditions, independent of registration or specific organ characteristics. Objective: This study aims to develop a robust framework that enhances the quality of ultrasound images, particularly those captured with compact, portable devices, which are often constrained by low quality due to hardware limitations. The framework is designed to effectively process nonregistered ultrasound image pairs, a common challenge in medical imaging, across various clinical settings and device types. By addressing these challenges, the research seeks to provide a more generalized and adaptable solution that can be widely applied across diverse medical scenarios, improving the accessibility and quality of diagnostic imaging. Methods: A retrospective analysis was conducted by using a cycle-consistent generative adversarial network (CycleGAN) framework enhanced with perceptual loss to improve the quality of ultrasound images, focusing on nonregistered image pairs from various organ systems. The perceptual loss was integrated to preserve anatomical integrity by comparing deep features extracted from pretrained neural networks. The model’s performance was evaluated against corresponding high-resolution images, ensuring that the enhanced outputs closely mimic those from high-end ultrasound devices. The model was trained and validated using a publicly available, diverse dataset to ensure robustness and generalizability across different imaging scenarios. Results: The advanced CycleGAN framework, enhanced with perceptual loss, significantly outperformed the previous state-of-the-art, stable CycleGAN, in multiple evaluation metrics. Specifically, our method achieved a structural similarity index of 0.2889 versus 0.2502 (P<.001), a peak signal-to-noise ratio of 15.8935 versus 14.9430 (P<.001), and a learned perceptual image patch similarity score of 0.4490 versus 0.5005 (P<.001). These results demonstrate the model’s superior ability to enhance image quality while preserving critical anatomical details, thereby improving diagnostic usefulness. Conclusions: This study presents a significant advancement in ultrasound imaging by leveraging a CycleGAN model enhanced with perceptual loss to bridge the quality gap between images from different devices. By processing nonregistered image pairs, the model not only enhances visual quality but also ensures the preservation of essential anatomical structures, crucial for accurate diagnosis. This approach holds the potential to democratize high-quality ultrasound imaging, making it accessible through low-cost portable devices, thereby improving health care outcomes, particularly in resource-limited settings. Future research will focus on further validation and optimization for clinical use.
- Open Access
- 27 March 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Enhancing Ultrasound Image Quality Across Disease Domains: Application of Cycle-Consistent Generative Adversarial Network and Perceptual Loss (Preprint)
BACKGROUND Numerous studies have explored image processing techniques aimed at enhancing ultrasound images to narrow the performance gap between low-quality portable devices and high-end ultrasound equipment. These investigations often use registered image pairs created by modifying the same image through methods like down sampling or adding noise, rather than using separate images from different machines. Additionally, they rely on organ-specific features, limiting the models’ generalizability across various imaging conditions and devices. The challenge remains to develop a universal framework capable of improving image quality across different devices and conditions, independent of registration or specific organ characteristics. OBJECTIVE This study aims to develop a robust framework that enhances the quality of ultrasound images, particularly those captured with compact, portable devices, which are often constrained by low quality due to hardware limitations. The framework is designed to effectively process nonregistered ultrasound image pairs, a common challenge in medical imaging, across various clinical settings and device types. By addressing these challenges, the research seeks to provide a more generalized and adaptable solution that can be widely applied across diverse medical scenarios, improving the accessibility and quality of diagnostic imaging. METHODS A retrospective analysis was conducted by using a cycle-consistent generative adversarial network (CycleGAN) framework enhanced with perceptual loss to improve the quality of ultrasound images, focusing on nonregistered image pairs from various organ systems. The perceptual loss was integrated to preserve anatomical integrity by comparing deep features extracted from pretrained neural networks. The model’s performance was evaluated against corresponding high-resolution images, ensuring that the enhanced outputs closely mimic those from high-end ultrasound devices. The model was trained and validated using a publicly available, diverse dataset to ensure robustness and generalizability across different imaging scenarios. RESULTS The advanced CycleGAN framework, enhanced with perceptual loss, significantly outperformed the previous state-of-the-art, stable CycleGAN, in multiple evaluation metrics. Specifically, our method achieved a structural similarity index of 0.2889 versus 0.2502 (P<.001), a peak signal-to-noise ratio of 15.8935 versus 14.9430 (P<.001), and a learned perceptual image patch similarity score of 0.4490 versus 0.5005 (P<.001). These results demonstrate the model’s superior ability to enhance image quality while preserving critical anatomical details, thereby improving diagnostic usefulness. CONCLUSIONS This study presents a significant advancement in ultrasound imaging by leveraging a CycleGAN model enhanced with perceptual loss to bridge the quality gap between images from different devices. By processing nonregistered image pairs, the model not only enhances visual quality but also ensures the preservation of essential anatomical structures, crucial for accurate diagnosis. This approach holds the potential to democratize high-quality ultrasound imaging, making it accessible through low-cost portable devices, thereby improving health care outcomes, particularly in resource-limited settings. Future research will focus on further validation and optimization for clinical use.
- 27 March 2024
- Published by JMIR Publications Inc. in JMIR Preprints
Open AccessEnhancing Ultrasound Image Quality Across Disease Domains: Application of Cycle-Consistent Generative Adversarial Network and Perceptual Loss (Preprint)
BACKGROUND Numerous studies have explored image processing techniques aimed at enhancing ultrasound images to narrow the performance gap between low-quality portable devices and high-end ultrasound equipment. These investigations often use registered image pairs created by modifying the same image through methods like down sampling or adding noise, rather than using separate images from different machines. Additionally, they rely on organ-specific features, limiting the models’ generalizability across various imaging conditions and devices. The challenge remains to develop a universal framework capable of improving image quality across different devices and conditions, independent of registration or specific organ characteristics. OBJECTIVE This study aims to develop a robust framework that enhances the quality of ultrasound images, particularly those captured with compact, portable devices, which are often constrained by low quality due to hardware limitations. The framework is designed to effectively process nonregistered ultrasound image pairs, a common challenge in medical imaging, across various clinical settings and device types. By addressing these challenges, the research seeks to provide a more generalized and adaptable solution that can be widely applied across diverse medical scenarios, improving the accessibility and quality of diagnostic imaging. METHODS A retrospective analysis was conducted by using a cycle-consistent generative adversarial network (CycleGAN) framework enhanced with perceptual loss to improve the quality of ultrasound images, focusing on nonregistered image pairs from various organ systems. The perceptual loss was integrated to preserve anatomical integrity by comparing deep features extracted from pretrained neural networks. The model’s performance was evaluated against corresponding high-resolution images, ensuring that the enhanced outputs closely mimic those from high-end ultrasound devices. The model was trained and validated using a publicly available, diverse dataset to ensure robustness and generalizability across different imaging scenarios. RESULTS The advanced CycleGAN framework, enhanced with perceptual loss, significantly outperformed the previous state-of-the-art, stable CycleGAN, in multiple evaluation metrics. Specifically, our method achieved a structural similarity index of 0.2889 versus 0.2502 (P<.001), a peak signal-to-noise ratio of 15.8935 versus 14.9430 (P<.001), and a learned perceptual image patch similarity score of 0.4490 versus 0.5005 (P<.001). These results demonstrate the model’s superior ability to enhance image quality while preserving critical anatomical details, thereby improving diagnostic usefulness. CONCLUSIONS This study presents a significant advancement in ultrasound imaging by leveraging a CycleGAN model enhanced with perceptual loss to bridge the quality gap between images from different devices. By processing nonregistered image pairs, the model not only enhances visual quality but also ensures the preservation of essential anatomical structures, crucial for accurate diagnosis. This approach holds the potential to democratize high-quality ultrasound imaging, making it accessible through low-cost portable devices, thereby improving health care outcomes, particularly in resource-limited settings. Future research will focus on further validation and optimization for clinical use.
- Open Access
- 22 March 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 30129-30145
On Learning Mechanical Laws of Motion From Video Using Neural Networks
In computer vision, physics plays an important role in several applications. In this work, we teach a machine to detect the mechanical laws of motion of physical objects using video, and show how the results can be useful for computer vision tasks. We assume no prior knowledge of physics, beyond a temporal stream of bounding boxes. The problem is very difficult because a machine must learn not only a governing equation (e.g. projectile motion) but also the existence of governing parameters (e.g. velocities). We evaluate our ability to represent the physical laws of motion in video, such as the movement of a projectile or circular motion, in both real and constructed videos. These elementary tasks have textbook governing equations and enable ground truth verification of our approach. To establish the importance of the proposed method, we show a real-world use case in the domain of object tracking in confounding scenes, where existing state-of-the-art algorithms fail. Incorporating physics into computer vision not only serves the purpose of curiosity-driven research, but also provides an inductive bias for computer vision applications like object tracking.
- 22 March 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 30129-30145
Open AccessOn Learning Mechanical Laws of Motion From Video Using Neural Networks
In computer vision, physics plays an important role in several applications. In this work, we teach a machine to detect the mechanical laws of motion of physical objects using video, and show how the results can be useful for computer vision tasks. We assume no prior knowledge of physics, beyond a temporal stream of bounding boxes. The problem is very difficult because a machine must learn not only a governing equation (e.g. projectile motion) but also the existence of governing parameters (e.g. velocities). We evaluate our ability to represent the physical laws of motion in video, such as the movement of a projectile or circular motion, in both real and constructed videos. These elementary tasks have textbook governing equations and enable ground truth verification of our approach. To establish the importance of the proposed method, we show a real-world use case in the domain of object tracking in confounding scenes, where existing state-of-the-art algorithms fail. Incorporating physics into computer vision not only serves the purpose of curiosity-driven research, but also provides an inductive bias for computer vision applications like object tracking.
- 3 November 2022
- Published by Springer Nature
- p. 326-343
MIME: Minority Inclusion for Majority Group Enhancement of AI Performance
- 3 November 2022
- Published by Springer Nature
- p. 326-343
MIME: Minority Inclusion for Majority Group Enhancement of AI Performance
- 1 July 2021
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 68
Screening of knee-joint vibroarthrographic signals using time and spectral domain features
- 1 July 2021
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 68
Screening of knee-joint vibroarthrographic signals using time and spectral domain features