Javad Alirezaie
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Javad Alirezaie has a total of 29 co-authors. Here are the top co-authors by number of shared publications:
- Paul Babyn
Department of Medical Imaging, University of Saskatchewan, Saskatoon, SK S7N 0W8, Canada
6 shared publications - Luella Marcos
Department of Electrical, Computer and Biomedical Engineering, Toronto Metropolitan University, Toronto, ON M5B 2K3, Canada
5 shared publications - Paul Babyn
Department of Medical Imaging, University of Saskatchewan and Saskatoon Health Region, Saskatoon, Canada
3 shared publications Farzan Niknejad Mazandarani
2 shared publications- Farzan Niknejad Mazandarani
Department of Electrical, Computer and Biomedical Engineering, Toronto Metropolitan University, Toronto, Canada
2 shared publications - Iqbal Banwait
Department of Electrical, Computer and Biomedical Engineering, Toronto Metropolitan University, Toronto, ON M5B 2K3, Canada
2 shared publications
Latest Publications
- Open Access
- 13 October 2026
- Published by Elsevier in SSRN Electronic Journal
URLSR: Designing Ultralight-weight Image Super Resolution Network using Reinforcement Learning
At an ultralight-weight budget of roughly 100K parameters, which neural operations a super-resolution network uses matters more than how its parameters are distributed. Yet, existing lightweight models repeat a single homogeneous block, and standard architecture search methods are not tailored to this compact regime. We propose URLSR, a reinforcement learning-based architecture search specialized for ultralight-weight super resolution. URLSR selects operations from a diverse action set spanning convolution, ConvLSTM, graph convolution, shifted-window attention, and state-space Mamba blocks. The parameter budget is structurally enforced, allowing the reward to optimize reconstruction fidelity directly alongside an entropy-based novelty term that prevents operator collapse. Furthermore, a learnable sub-pixel translation module aligns feature spectra in the frequency domain, recovering fine geometric textures without interpolation blur. Evaluated against ultralight-weight variants of six state-of-the-art networks across five benchmarks, URLSR ranks first in 27 of 30 settings at a one-time search cost of 23 GPU-hours.
- 13 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessURLSR: Designing Ultralight-weight Image Super Resolution Network using Reinforcement Learning
At an ultralight-weight budget of roughly 100K parameters, which neural operations a super-resolution network uses matters more than how its parameters are distributed. Yet, existing lightweight models repeat a single homogeneous block, and standard architecture search methods are not tailored to this compact regime. We propose URLSR, a reinforcement learning-based architecture search specialized for ultralight-weight super resolution. URLSR selects operations from a diverse action set spanning convolution, ConvLSTM, graph convolution, shifted-window attention, and state-space Mamba blocks. The parameter budget is structurally enforced, allowing the reward to optimize reconstruction fidelity directly alongside an entropy-based novelty term that prevents operator collapse. Furthermore, a learnable sub-pixel translation module aligns feature spectra in the frequency domain, recovering fine geometric textures without interpolation blur. Evaluated against ultralight-weight variants of six state-of-the-art networks across five benchmarks, URLSR ranks first in 27 of 30 settings at a one-time search cost of 23 GPU-hours.
- Open Access
SSRaDNet: An Empirical Study of Radar Object Detection Evaluation and Deployment Trade-Offs
Millimeter-wave radar is an attractive sensing modality for autonomous driving because of its robustness under adverse weather and lighting conditions. However, comparisons between radar object detectors are hindered by inconsistent evaluation procedures and limited reporting of deployment costs. This work presents a comprehensive benchmark using standardized VOC mean average precision (mAP) and evaluates processing-pipeline latency across preprocessing, network inference, and post-processing. We evaluate SSRaDNet, a lightweight hybrid CNN–Swin Transformer detector, across analog-to-digital converter (ADC), range-Doppler (RD), and range-azimuth-Doppler (RAD) representations and compare it with existing methods on RADIal and RADDet. SSRaDNet achieves state-of-the-art AP@0.5 among re-evaluated RADIal models using RD input, reaching 88.89%, and provides the strongest overall balance of detection and segmentation performance among the compared ADC configurations. On RADDet, it achieves the best performance among methods using the native-resolution RAD tensor (45.69% mAP@0.5) and introduces the first ADC-based detector with full bounding-box regression. We also evaluate alternative post-processing strategies, including hybrid MaxPoolNMS–GreedyNMS, to characterize accuracy–latency trade-offs associated with alternative post-processing strategies. These results demonstrate the importance of standardized evaluation and full-pipeline runtime analysis for radar-based object detection.
Open AccessSSRaDNet: An Empirical Study of Radar Object Detection Evaluation and Deployment Trade-Offs
Millimeter-wave radar is an attractive sensing modality for autonomous driving because of its robustness under adverse weather and lighting conditions. However, comparisons between radar object detectors are hindered by inconsistent evaluation procedures and limited reporting of deployment costs. This work presents a comprehensive benchmark using standardized VOC mean average precision (mAP) and evaluates processing-pipeline latency across preprocessing, network inference, and post-processing. We evaluate SSRaDNet, a lightweight hybrid CNN–Swin Transformer detector, across analog-to-digital converter (ADC), range-Doppler (RD), and range-azimuth-Doppler (RAD) representations and compare it with existing methods on RADIal and RADDet. SSRaDNet achieves state-of-the-art AP@0.5 among re-evaluated RADIal models using RD input, reaching 88.89%, and provides the strongest overall balance of detection and segmentation performance among the compared ADC configurations. On RADDet, it achieves the best performance among methods using the native-resolution RAD tensor (45.69% mAP@0.5) and introduces the first ADC-based detector with full bounding-box regression. We also evaluate alternative post-processing strategies, including hybrid MaxPoolNMS–GreedyNMS, to characterize accuracy–latency trade-offs associated with alternative post-processing strategies. These results demonstrate the importance of standardized evaluation and full-pipeline runtime analysis for radar-based object detection.
- Open Access
Physics-Guided Sequential State Space Transformer (PS3T) for Projection Domain LDCT Denoising
Low-Dose Computed Tomography (LDCT) reduces radiation exposure but introduces severe quantum noise and streak artifacts that degrade image quality. To address these challenges, we propose the Physics-Guided Sequential State Space Transformer (), a projection-domain denoising framework that combines sequential state-space modeling with a photon-aware attention mechanism to capture long-range dependencies across projection angles with linear computational complexity. A differentiable Filtered Backprojection (FBP) layer further enforces reconstruction-domain consistency during training. The proposed framework was evaluated on the Mayo Clinic LDCT and Projection Dataset using patient-level dataset partitioning. Experimental results demonstrate that consistently outperforms state-of-the-art methods, including DRL, SADiff, and GEDFormer, across the abdomen, head, and chest datasets. On the abdomen dataset, reached a peak PSNR of 42.40 dB, an SSIM of 0.9020, and the lowest RMSE of 0.0076 across anatomical regions. Statistical analysis using 95% confidence intervals and paired Wilcoxon signed-rank tests confirmed that these improvements were significant (). Furthermore, achieved the lowest reconstruction consistency loss (0.0128 at epoch 50), demonstrating stable convergence and the effectiveness of incorporating acquisition physics into projection-domain LDCT denoising.Open AccessPhysics-Guided Sequential State Space Transformer (PS3T) for Projection Domain LDCT Denoising
Low-Dose Computed Tomography (LDCT) reduces radiation exposure but introduces severe quantum noise and streak artifacts that degrade image quality. To address these challenges, we propose the Physics-Guided Sequential State Space Transformer (), a projection-domain denoising framework that combines sequential state-space modeling with a photon-aware attention mechanism to capture long-range dependencies across projection angles with linear computational complexity. A differentiable Filtered Backprojection (FBP) layer further enforces reconstruction-domain consistency during training. The proposed framework was evaluated on the Mayo Clinic LDCT and Projection Dataset using patient-level dataset partitioning. Experimental results demonstrate that consistently outperforms state-of-the-art methods, including DRL, SADiff, and GEDFormer, across the abdomen, head, and chest datasets. On the abdomen dataset, reached a peak PSNR of 42.40 dB, an SSIM of 0.9020, and the lowest RMSE of 0.0076 across anatomical regions. Statistical analysis using 95% confidence intervals and paired Wilcoxon signed-rank tests confirmed that these improvements were significant (). Furthermore, achieved the lowest reconstruction consistency loss (0.0128 at epoch 50), demonstrating stable convergence and the effectiveness of incorporating acquisition physics into projection-domain LDCT denoising. - Open Access
- 19 August 2026
- Published by MDPI AG in Preprints.org
SSRaDNet: An Empirical Study of Radar Object Detection Evaluation and Deployment Trade-Offs
Millimeter-wave radar is an attractive sensing modality for autonomous driving because of its robustness under adverse weather and lighting conditions. However, comparisons between radar object detectors are hindered by inconsistent evaluation procedures and limited reporting of deployment costs. This work presents a comprehensive benchmark using standardized VOC mean average precision (mAP) and evaluates end-to-end latency across preprocessing, network inference, and postprocessing. We evaluate SSRaDNet, a lightweight hybrid CNN–Swin Transformer detector, across analog-to-digital converter (ADC), range-Doppler (RD), and range-azimuth-Doppler (RAD) representations and compare it with existing methods on RADIal and RADDet. SSRaDNet achieves state-of-the-art AP@0.5 among re-evaluated RADIal models using RD input, reaching 88.89%, and provides the strongest overall balance of detection and segmentation performance among the compared ADC configurations. On RADDet, it achieves the best performance among methods using the native-resolution RAD tensor (45.69% mAP@0.5) and introduces the first ADC-based detector with full bounding-box regression. We also evaluate alternative postprocessing strategies, including hybrid MaxPoolNMS-GreedyNMS, to characterize deployment trade-offs on resource-constrained hardware. These results demonstrate the importance of standardized evaluation and full-pipeline runtime analysis for radar-based object detection. Code is available at https://github.com/IqbalBan/SSRadNet.
- 19 August 2026
- Published by MDPI AG in Preprints.org
Open AccessSSRaDNet: An Empirical Study of Radar Object Detection Evaluation and Deployment Trade-Offs
Millimeter-wave radar is an attractive sensing modality for autonomous driving because of its robustness under adverse weather and lighting conditions. However, comparisons between radar object detectors are hindered by inconsistent evaluation procedures and limited reporting of deployment costs. This work presents a comprehensive benchmark using standardized VOC mean average precision (mAP) and evaluates end-to-end latency across preprocessing, network inference, and postprocessing. We evaluate SSRaDNet, a lightweight hybrid CNN–Swin Transformer detector, across analog-to-digital converter (ADC), range-Doppler (RD), and range-azimuth-Doppler (RAD) representations and compare it with existing methods on RADIal and RADDet. SSRaDNet achieves state-of-the-art AP@0.5 among re-evaluated RADIal models using RD input, reaching 88.89%, and provides the strongest overall balance of detection and segmentation performance among the compared ADC configurations. On RADDet, it achieves the best performance among methods using the native-resolution RAD tensor (45.69% mAP@0.5) and introduces the first ADC-based detector with full bounding-box regression. We also evaluate alternative postprocessing strategies, including hybrid MaxPoolNMS-GreedyNMS, to characterize deployment trade-offs on resource-constrained hardware. These results demonstrate the importance of standardized evaluation and full-pipeline runtime analysis for radar-based object detection. Code is available at https://github.com/IqbalBan/SSRadNet.
- Open Access
- 1 October 2025
- Published by Wiley in International Journal of Imaging Systems and Technology
- Vol. 35 (6)
Low‐Dose Computed Tomography Image Denoising Vision Transformer Model Optimization Using Space State Method
Low‐dose computed tomography (LDCT) is widely used to promote reduction of patient radiation exposure, but the associated increase in image noise poses challenges for diagnostic accuracy. In this study, we propose a Vision Transformer (ViT)‐based denoising framework enhanced with a State Space Optimizing Block (SSOB) to improve both image quality and computational efficiency. The SSOB upgrades the multihead self‐attention mechanism by reducing spatial redundancy and optimizing contextual feature fusion, thereby strengthening the transformer's ability to capture long‐range dependencies and preserve fine anatomical structures under severe noise. Extensive evaluations on randomized and categorized datasets demonstrate that the proposed model consistently outperforms existing state‐of‐the‐art denoising approaches. It achieved the highest average SSIM (up to 6.10% improvement), PSNR values (36.51 ± 0.37 dB on randomized and 36.30 ± 0.36 dB on categorized datasets), and the lowest RMSE, surpassing recent CNN‐transformer‐based denoising hybrid models by approximately 12%. Intensity profile analysis further confirmed its effectiveness, showing sharper edge transitions and more accurate gray‐level distributions across anatomical boundaries, closely aligning with ground truth and retaining subtle diagnostic features often lost in competing models. In addition to improved reconstruction quality, the SSOB‐empowered ViT achieved notable computational gains. It delivered the fastest inference (0.42 s per image), highest throughput (2.38 images/s), lowest GPU memory usage (750 MB), and smallest model size (7.6 MB), alongside one of the shortest training times (6.5 h). Compared to legacy architectures, which required up to 16 h of training and substantially more resources, the proposed model offers both accuracy and deployability. Collectively, these findings establish the SSOB as a key component for efficient transformer‐based LDCT denoising, addressing memory and convergence challenges while preserving global contextual advantages.
- 1 October 2025
- Published by Wiley in International Journal of Imaging Systems and Technology
- Vol. 35 (6)
Open AccessLow‐Dose Computed Tomography Image Denoising Vision Transformer Model Optimization Using Space State Method
Low‐dose computed tomography (LDCT) is widely used to promote reduction of patient radiation exposure, but the associated increase in image noise poses challenges for diagnostic accuracy. In this study, we propose a Vision Transformer (ViT)‐based denoising framework enhanced with a State Space Optimizing Block (SSOB) to improve both image quality and computational efficiency. The SSOB upgrades the multihead self‐attention mechanism by reducing spatial redundancy and optimizing contextual feature fusion, thereby strengthening the transformer's ability to capture long‐range dependencies and preserve fine anatomical structures under severe noise. Extensive evaluations on randomized and categorized datasets demonstrate that the proposed model consistently outperforms existing state‐of‐the‐art denoising approaches. It achieved the highest average SSIM (up to 6.10% improvement), PSNR values (36.51 ± 0.37 dB on randomized and 36.30 ± 0.36 dB on categorized datasets), and the lowest RMSE, surpassing recent CNN‐transformer‐based denoising hybrid models by approximately 12%. Intensity profile analysis further confirmed its effectiveness, showing sharper edge transitions and more accurate gray‐level distributions across anatomical boundaries, closely aligning with ground truth and retaining subtle diagnostic features often lost in competing models. In addition to improved reconstruction quality, the SSOB‐empowered ViT achieved notable computational gains. It delivered the fastest inference (0.42 s per image), highest throughput (2.38 images/s), lowest GPU memory usage (750 MB), and smallest model size (7.6 MB), alongside one of the shortest training times (6.5 h). Compared to legacy architectures, which required up to 16 h of training and substantially more resources, the proposed model offers both accuracy and deployability. Collectively, these findings establish the SSOB as a key component for efficient transformer‐based LDCT denoising, addressing memory and convergence challenges while preserving global contextual advantages.
- Open Access
- 31 July 2025
- Published by Elsevier in Computer Methods and Programs in Biomedicine Update
- Vol. 8
Robust lung segmentation in Chest X-ray images using modified U-Net with deeper network and residual blocks
- 31 July 2025
- Published by Elsevier in Computer Methods and Programs in Biomedicine Update
- Vol. 8
Open AccessRobust lung segmentation in Chest X-ray images using modified U-Net with deeper network and residual blocks
- Open Access
- 17 March 2025
- Published by Wiley in International Journal of Imaging Systems and Technology
- Vol. 35 (2)
A Time‐Adaptive Diffusion‐Based CT Image Denoising Method by Processing Directional and Non‐Local Information
Low‐dose computed tomography (CT) images are prone to noise and artifacts caused by photon starvation and electronic noise. Recently, researchers have explored the use of transformer‐based neural networks combined with generative diffusion models, showing promising results in denoising CT images. Despite their high performance, these approaches often struggle to process crucial information in the input data, resulting in suboptimal image quality. To address this limitation, we propose Starformer, a novel transformer‐based operation designed to extract non‐local directional features essential for diagnostic accuracy while maintaining an acceptable computational complexity overhead. Starformer is seamlessly integrated into the time‐adaptive schedules of a diffusion model, dynamically balancing global structural extraction and fine texture refinement throughout the diffusion process. This enables the generation of high‐quality, realistic textures in the final denoised images. Extensive experimental results demonstrate the effectiveness of both approaches in enhancing CT image quality, with improvements of up to 15% in PSNR and 36% in SSIM, highlighting their superiority over state‐of‐the‐art methods.
- 17 March 2025
- Published by Wiley in International Journal of Imaging Systems and Technology
- Vol. 35 (2)
Open AccessA Time‐Adaptive Diffusion‐Based CT Image Denoising Method by Processing Directional and Non‐Local Information
Low‐dose computed tomography (CT) images are prone to noise and artifacts caused by photon starvation and electronic noise. Recently, researchers have explored the use of transformer‐based neural networks combined with generative diffusion models, showing promising results in denoising CT images. Despite their high performance, these approaches often struggle to process crucial information in the input data, resulting in suboptimal image quality. To address this limitation, we propose Starformer, a novel transformer‐based operation designed to extract non‐local directional features essential for diagnostic accuracy while maintaining an acceptable computational complexity overhead. Starformer is seamlessly integrated into the time‐adaptive schedules of a diffusion model, dynamically balancing global structural extraction and fine texture refinement throughout the diffusion process. This enables the generation of high‐quality, realistic textures in the final denoised images. Extensive experimental results demonstrate the effectiveness of both approaches in enhancing CT image quality, with improvements of up to 15% in PSNR and 36% in SSIM, highlighting their superiority over state‐of‐the‐art methods.
- 10 March 2025
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 38 (6) , 4255-4275
SADiff: A Sinogram-Aware Diffusion Model for Low-Dose CT Image Denoising
CT image denoising is a crucial task in medical imaging systems, aimed at enhancing the quality of acquired visual signals. The emergence of diffusion models in machine learning has revolutionized the generation of high-quality CT images. However, diffusion-based CT image denoising methods suffer from two key shortcomings. First, they do not incorporate image formation priors from CT imaging, which limits their adaptability to the CT image denoising task. Second, they are trained on CT images with varying structures and textures at the signal phase, which hinders the model generalization capability. To address the first limitation, we propose a novel conditioning module for our diffusion model that leverages image formation priors from the sinogram domain to generate rich features. To tackle the second issue, we introduce a two-phase training mechanism in which the network gradually learns different anatomical textures and structures. Extensive experimental results demonstrate the effectiveness of both approaches in enhancing CT image quality, with improvements of up to 17% in PSNR and 38% in SSIM, highlighting their superiority over state-of-the-art methods.
- 10 March 2025
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 38 (6) , 4255-4275
SADiff: A Sinogram-Aware Diffusion Model for Low-Dose CT Image Denoising
CT image denoising is a crucial task in medical imaging systems, aimed at enhancing the quality of acquired visual signals. The emergence of diffusion models in machine learning has revolutionized the generation of high-quality CT images. However, diffusion-based CT image denoising methods suffer from two key shortcomings. First, they do not incorporate image formation priors from CT imaging, which limits their adaptability to the CT image denoising task. Second, they are trained on CT images with varying structures and textures at the signal phase, which hinders the model generalization capability. To address the first limitation, we propose a novel conditioning module for our diffusion model that leverages image formation priors from the sinogram domain to generate rich features. To tackle the second issue, we introduce a two-phase training mechanism in which the network gradually learns different anatomical textures and structures. Extensive experimental results demonstrate the effectiveness of both approaches in enhancing CT image quality, with improvements of up to 17% in PSNR and 38% in SSIM, highlighting their superiority over state-of-the-art methods.
- Open Access
- 3 March 2025
- Published by MDPI AG in Algorithms
- Vol. 18 (3) , 134
Edge Detection Attention Module in Pure Vision Transformer for Low-Dose X-Ray Computed Tomography Image Denoising
X-ray computed tomography (CT) is vital for medical diagnostics, but frequent radiation exposure raises concerns, driving the adoption of low-dose CT (LDCT) to mitigate risks. However, LDCT often introduces noise, compromising diagnostic accuracy. This paper proposes a pure vision transformer (PViT) for LDCT denoising, enhanced with a gradient–Laplacian attention module (GLAM) to improve edge preservation and fine structural detail reconstruction. The model’s robustness was validated across five diverse datasets (piglet, head, abdomen, chest, thoracic), demonstrating consistent performance in preserving anatomical structures. Extensive ablation studies on attention configurations and loss functions further substantiated the contributions of each module. Quantitative evaluation using PSNR and SSIM, alongside radiologist assessment, confirmed significant noise suppression and sharper anatomical boundaries, particularly in regions with fine details such as organ interfaces and bone structures. Additionally, in benchmark comparisons against state-of-the-art LDCT models (RED-CNN, TED-Net, DSC-GAN, DRL-EMP) and traditional methods (BM3D), the model exhibited lower parameter and stable training performance. These findings highlight the model’s robustness, efficiency, and clinical applicability, making it a promising solution for improving LDCT image quality while maintaining computational efficiency.
- 3 March 2025
- Published by MDPI AG in Algorithms
- Vol. 18 (3) , 134
Open AccessEdge Detection Attention Module in Pure Vision Transformer for Low-Dose X-Ray Computed Tomography Image Denoising
X-ray computed tomography (CT) is vital for medical diagnostics, but frequent radiation exposure raises concerns, driving the adoption of low-dose CT (LDCT) to mitigate risks. However, LDCT often introduces noise, compromising diagnostic accuracy. This paper proposes a pure vision transformer (PViT) for LDCT denoising, enhanced with a gradient–Laplacian attention module (GLAM) to improve edge preservation and fine structural detail reconstruction. The model’s robustness was validated across five diverse datasets (piglet, head, abdomen, chest, thoracic), demonstrating consistent performance in preserving anatomical structures. Extensive ablation studies on attention configurations and loss functions further substantiated the contributions of each module. Quantitative evaluation using PSNR and SSIM, alongside radiologist assessment, confirmed significant noise suppression and sharper anatomical boundaries, particularly in regions with fine details such as organ interfaces and bone structures. Additionally, in benchmark comparisons against state-of-the-art LDCT models (RED-CNN, TED-Net, DSC-GAN, DRL-EMP) and traditional methods (BM3D), the model exhibited lower parameter and stable training performance. These findings highlight the model’s robustness, efficiency, and clinical applicability, making it a promising solution for improving LDCT image quality while maintaining computational efficiency.