2026
9.
Gedik, Hakan Emre; Martin, Andrew; Munir, Mustafa; Baser, Oguzhan; Marculescu, Radu; Chinchali, Sandeep P.; Bovik, Alan C.
AttentionViG: Cross-Attention-Based Dynamic Neighbor Aggregation in Vision GNNs Conference Forthcoming
The Fifth Learning on Graphs Conference, Forthcoming.
Links | BibTeX | Tags: Deep Learning, Deep Learning Architecture, Efficient AI, Graph Neural Network
@conference{AttentionViG,
title = {AttentionViG: Cross-Attention-Based Dynamic Neighbor Aggregation in Vision GNNs},
author = {Hakan Emre Gedik and Andrew Martin and Mustafa Munir and Oguzhan Baser and Radu Marculescu and Sandeep P. Chinchali and Alan C. Bovik},
url = {https://www.arxiv.org/abs/2509.25570},
year = {2026},
date = {2026-09-26},
urldate = {2025-09-29},
booktitle = {The Fifth Learning on Graphs Conference},
keywords = {Deep Learning, Deep Learning Architecture, Efficient AI, Graph Neural Network},
pubstate = {forthcoming},
tppubtype = {conference}
}
8.
Munir, Mustafa; Zalewski, Sophia; Liu, Shiqiu; Tarjan, David; Belede, Sushmitha; Patney, Anjul; Marculescu, Radu
SmoothDiffusion-VE: Real-time Generative Video Editing Using Adaptive Feature Cache Conference
2026 Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026), 2026.
Links | BibTeX | Tags: Edge AI, Efficient AI, Efficient Inference, Generative AI
@conference{SmoothDiffusion_WACV_2026,
title = {SmoothDiffusion-VE: Real-time Generative Video Editing Using Adaptive Feature Cache},
author = {Mustafa Munir and Sophia Zalewski and Shiqiu Liu and David Tarjan and Sushmitha Belede and Anjul Patney and Radu Marculescu},
url = {https://openaccess.thecvf.com/content/WACV2026/html/Munir_SmoothDiffusion-VE_Real-time_Generative_Video_Editing_Using_Adaptive_Feature_Cache_WACV_2026_paper.html},
year = {2026},
date = {2026-03-02},
urldate = {2026-03-02},
booktitle = {2026 Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026)},
keywords = {Edge AI, Efficient AI, Efficient Inference, Generative AI},
pubstate = {published},
tppubtype = {conference}
}
7.
Munir, Mustafa; Rahman, Md Mostafijur; Marculescu, Radu
AdaptViG: Adaptive Vision GNN with Exponential Decay Gating Conference
2026 Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026), 2026.
Links | BibTeX | Tags: Deep Learning, Deep Learning Architecture, Efficient AI, Featured, Graph Neural Network
@conference{AdaptViG_WACV_2026,
title = {AdaptViG: Adaptive Vision GNN with Exponential Decay Gating},
author = {Mustafa Munir and Md Mostafijur Rahman and Radu Marculescu},
url = {https://ieeexplore.ieee.org/abstract/document/11492666},
year = {2026},
date = {2026-03-02},
urldate = {2026-03-02},
booktitle = {2026 Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026)},
keywords = {Deep Learning, Deep Learning Architecture, Efficient AI, Featured, Graph Neural Network},
pubstate = {published},
tppubtype = {conference}
}
6.
Shah, Sahil; Sharan, S P; Goel, Harsh; Choi, Minkyu; Munir, Mustafa; Pasula, Manvik; Marculescu, Radu; Chinchali, Sandeep
NeuS-QA: Grounding Long-Form Video Understanding in Temporal Logic and Neuro-Symbolic Reasoning Conference
AAAI Conference on Artificial Intelligence 2026, vol. 40, no. 11, 2026.
Links | BibTeX | Tags: Deep Learning, Efficient AI, Featured, Generative AI
@conference{NeuS-QA,
title = {NeuS-QA: Grounding Long-Form Video Understanding in Temporal Logic and Neuro-Symbolic Reasoning},
author = {Sahil Shah and S P Sharan and Harsh Goel and Minkyu Choi and Mustafa Munir and Manvik Pasula and Radu Marculescu and Sandeep Chinchali},
url = {https://ojs.aaai.org/index.php/AAAI/article/view/37834},
year = {2026},
date = {2026-01-15},
urldate = {2026-01-15},
booktitle = {AAAI Conference on Artificial Intelligence 2026},
volume = {40},
number = {11},
pages = {8805\textendash8813},
keywords = {Deep Learning, Efficient AI, Featured, Generative AI},
pubstate = {published},
tppubtype = {conference}
}
2025
5.
Munir, Mustafa; Rahman, Md Mostafijur; Wei, Xiwen; Yang, Yuedong; Marculescu, Radu
SearchViG: Optimal Vision GNNs via Ramanujan Spectral Optimization Conference
The Fourth Learning on Graphs Conference (LOG 2025), 2025.
Links | BibTeX | Tags: Deep Learning, Deep Learning Architecture, Dynamic networks, Efficient AI, Featured, Graph Neural Network
@conference{SearchViG_LOG_2025,
title = {SearchViG: Optimal Vision GNNs via Ramanujan Spectral Optimization},
author = {Mustafa Munir and Md Mostafijur Rahman and Xiwen Wei and Yuedong Yang and Radu Marculescu},
url = {https://openreview.net/pdf?id=cmEzgaYIJC},
year = {2025},
date = {2025-12-15},
urldate = {2025-12-15},
booktitle = {The Fourth Learning on Graphs Conference (LOG 2025)},
keywords = {Deep Learning, Deep Learning Architecture, Dynamic networks, Efficient AI, Featured, Graph Neural Network},
pubstate = {published},
tppubtype = {conference}
}
4.
Rahman, Md Mostafijur; Marculescu, Radu
MK-UNet: Multi-kernel Lightweight CNN for Medical Image Segmentation Conference
Proceedings of the International Conference on Computer Vision (ICCV) Workshops, 2025.
Abstract | Links | BibTeX | Tags: Efficient AI, Lightweight Architecture, Medical Image Segmentation, Multi-kernel Convolutions
@conference{mkunet@rahman,
title = {MK-UNet: Multi-kernel Lightweight CNN for Medical Image Segmentation},
author = {Md Mostafijur Rahman and Radu Marculescu},
url = {https://openaccess.thecvf.com/content/ICCV2025W/CVAMD/html/Rahman_MK-UNet_Multi-kernel_Lightweight_CNN_for_Medical_Image_Segmentation_ICCVW_2025_paper.html},
year = {2025},
date = {2025-10-19},
urldate = {2025-10-19},
publisher = {Proceedings of the International Conference on Computer Vision (ICCV) Workshops},
abstract = {In this paper, we introduce MK-UNet, a paradigm shift towards ultra-lightweight, multi-kernel U-shaped CNNs tailored for medical image segmentation. Central to MK-UNet is the multi-kernel depth-wise convolution block (MKDC) we design to adeptly process images through multiple kernels, while capturing complex multi-resolution spatial relationships. MK-UNet also emphasizes the images salient features through sophisticated attention mechanisms, including channel, spatial, and grouped gated attention. Our MK-UNet network, with a modest computational footprint of only 0.316 M parameters and 0.314 G FLOPs, represents not only a remarkably lightweight, but also significantly improved segmentation solution that provides higher accuracy over state-of-the-art (SOTA) methods across six binary medical imaging benchmarks. Specifically, MK-UNet outperforms TransUNet in DICE score with nearly 333x and 123x fewer parameters and FLOPs, respectively. Similarly, when compared against UNeXt, MK-UNet exhibits superior segmentation performance, improving the DICE score up to 6.7% margins while operating with 4.7 x fewer# Params. Our MK-UNet also outperforms other recent lightweight networks, such as MedT, CMUNeXt, EGE-UNet, and Rolling-UNet, with much lower computational resources. This leap in performance, coupled with drastic computational gains, positions MK-UNet as an unparalleled solution for real-time, high-fidelity medical diagnostics in resource-limited settings, such as point-of-care devices. Our implementation is available at https://github. com/SLDGroup/MK-UNet.},
keywords = {Efficient AI, Lightweight Architecture, Medical Image Segmentation, Multi-kernel Convolutions},
pubstate = {published},
tppubtype = {conference}
}
In this paper, we introduce MK-UNet, a paradigm shift towards ultra-lightweight, multi-kernel U-shaped CNNs tailored for medical image segmentation. Central to MK-UNet is the multi-kernel depth-wise convolution block (MKDC) we design to adeptly process images through multiple kernels, while capturing complex multi-resolution spatial relationships. MK-UNet also emphasizes the images salient features through sophisticated attention mechanisms, including channel, spatial, and grouped gated attention. Our MK-UNet network, with a modest computational footprint of only 0.316 M parameters and 0.314 G FLOPs, represents not only a remarkably lightweight, but also significantly improved segmentation solution that provides higher accuracy over state-of-the-art (SOTA) methods across six binary medical imaging benchmarks. Specifically, MK-UNet outperforms TransUNet in DICE score with nearly 333x and 123x fewer parameters and FLOPs, respectively. Similarly, when compared against UNeXt, MK-UNet exhibits superior segmentation performance, improving the DICE score up to 6.7% margins while operating with 4.7 x fewer# Params. Our MK-UNet also outperforms other recent lightweight networks, such as MedT, CMUNeXt, EGE-UNet, and Rolling-UNet, with much lower computational resources. This leap in performance, coupled with drastic computational gains, positions MK-UNet as an unparalleled solution for real-time, high-fidelity medical diagnostics in resource-limited settings, such as point-of-care devices. Our implementation is available at https://github. com/SLDGroup/MK-UNet.
3.
Munir, Mustafa; Zhang, Alex; Marculescu, Radu
VCMamba: Bridging Convolutions with Multi-Directional Mamba for Efficient Visual Representation Conference
Proceedings of the International Conference on Computer Vision (ICCV 2025) Workshops, 2025.
Abstract | Links | BibTeX | Tags: Deep Learning Architecture, Efficient AI, Featured, Lightweight Architecture, Model Compression & Optimization
@conference{VCMamba_ICCV_2025,
title = {VCMamba: Bridging Convolutions with Multi-Directional Mamba for Efficient Visual Representation},
author = {Mustafa Munir and Alex Zhang and Radu Marculescu},
url = {https://arxiv.org/abs/2509.04669},
year = {2025},
date = {2025-10-19},
urldate = {2025-10-19},
publisher = {Proceedings of the International Conference on Computer Vision (ICCV 2025) Workshops},
abstract = {Recent advances in Vision Transformers (ViTs) and State Space Models (SSMs) have challenged the dominance of Convolutional Neural Networks (CNNs) in computer vision. ViTs excel at capturing global context, and SSMs like Mamba offer linear complexity for long sequences, yet they do not capture fine-grained local features as effectively as CNNs. Conversely, CNNs possess strong inductive biases for local features but lack the global reasoning capabilities of transformers and Mamba. To bridge this gap, we introduce VCMamba, a novel vision backbone that integrates the strengths of CNNs and multi-directional Mamba SSMs. VCMamba employs a convolutional stem and a hierarchical structure with convolutional blocks in its early stages to extract rich local features. These convolutional blocks are then processed by later stages incorporating multi-directional Mamba blocks designed to efficiently model long-range dependencies and global context. This hybrid design allows for superior feature representation while maintaining linear complexity with respect to image resolution. We demonstrate VCMamba’s effectiveness through extensive experiments on ImageNet-1K classification and ADE20K semantic segmentation. Our VCMamba-B achieves 82.6% top-1 accuracy on ImageNet-1K, surpassing PlainMamba-L3 by 0.3% with 37% fewer parameters, and outperforming Vision GNN-B by 0.3% with 64% fewer parameters. Furthermore, VCMamba-B obtains 47.1 mIoU on ADE20K, exceeding EfficientFormer-L7 by 2.0 mIoU while utilizing 62% fewer parameters. Code is available at https://github.com/Wertyuui345/VCMamba.},
keywords = {Deep Learning Architecture, Efficient AI, Featured, Lightweight Architecture, Model Compression \& Optimization},
pubstate = {published},
tppubtype = {conference}
}
Recent advances in Vision Transformers (ViTs) and State Space Models (SSMs) have challenged the dominance of Convolutional Neural Networks (CNNs) in computer vision. ViTs excel at capturing global context, and SSMs like Mamba offer linear complexity for long sequences, yet they do not capture fine-grained local features as effectively as CNNs. Conversely, CNNs possess strong inductive biases for local features but lack the global reasoning capabilities of transformers and Mamba. To bridge this gap, we introduce VCMamba, a novel vision backbone that integrates the strengths of CNNs and multi-directional Mamba SSMs. VCMamba employs a convolutional stem and a hierarchical structure with convolutional blocks in its early stages to extract rich local features. These convolutional blocks are then processed by later stages incorporating multi-directional Mamba blocks designed to efficiently model long-range dependencies and global context. This hybrid design allows for superior feature representation while maintaining linear complexity with respect to image resolution. We demonstrate VCMamba’s effectiveness through extensive experiments on ImageNet-1K classification and ADE20K semantic segmentation. Our VCMamba-B achieves 82.6% top-1 accuracy on ImageNet-1K, surpassing PlainMamba-L3 by 0.3% with 37% fewer parameters, and outperforming Vision GNN-B by 0.3% with 64% fewer parameters. Furthermore, VCMamba-B obtains 47.1 mIoU on ADE20K, exceeding EfficientFormer-L7 by 2.0 mIoU while utilizing 62% fewer parameters. Code is available at https://github.com/Wertyuui345/VCMamba.
2.
Rahman, Md Mostafijur; Munir, Mustafa; Marculescu, Radu
EfficientMedNeXt: Multi-Receptive Dilated Convolutions for Medical Image Segmentation Conference
International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2025.
Abstract | Links | BibTeX | Tags: 3D Segmentation, Deep Learning Architecture, Efficient AI, Efficient Inference, Featured, Medical Image Segmentation
@conference{efficientmednext@rahman,
title = {EfficientMedNeXt: Multi-Receptive Dilated Convolutions for Medical Image Segmentation},
author = {Md Mostafijur Rahman and Mustafa Munir and Radu Marculescu},
url = {https://link.springer.com/chapter/10.1007/978-3-032-04965-0_19},
year = {2025},
date = {2025-09-23},
urldate = {2025-09-23},
publisher = {International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI)},
abstract = {In this work, we introduce EfficientMedNeXt\textemdasha lightweight, high-performance segmentation architecture developed through a two-phase optimization process applied to the MedNeXt architecture. To this end, we first optimize the decoder by reducing the high-resolution redundancy and unifying the decoder channels across stages for enhanced efficiency. Then, we introduce a new Dilated Multi-Receptive Field Block (DMRFB) to capture the multi-scale spatial context efficiently without increasing the kernel sizes and relying on the channel expansion convolutions. Extensive evaluations on BTCV, FeTA, and MSD show that EfficientMedNeXt-L achieves 87.0% DICE score on BTCV (+1.04% over MedNeXt-L) with 96.5% fewer parameters and 77.03% lower FLOPs. In addition, EfficientMedNeXt-S offers comparable DICE score, improved HD95, and 78.1% higher throughput while reducing parameters by 98.5% and FLOPs by 95%. These results demonstrate EfficientMedNeXt’s efficiency and accuracy, making it well-suited for real-world clinical applications. Code will be released upon acceptance.},
keywords = {3D Segmentation, Deep Learning Architecture, Efficient AI, Efficient Inference, Featured, Medical Image Segmentation},
pubstate = {published},
tppubtype = {conference}
}
In this work, we introduce EfficientMedNeXt—a lightweight, high-performance segmentation architecture developed through a two-phase optimization process applied to the MedNeXt architecture. To this end, we first optimize the decoder by reducing the high-resolution redundancy and unifying the decoder channels across stages for enhanced efficiency. Then, we introduce a new Dilated Multi-Receptive Field Block (DMRFB) to capture the multi-scale spatial context efficiently without increasing the kernel sizes and relying on the channel expansion convolutions. Extensive evaluations on BTCV, FeTA, and MSD show that EfficientMedNeXt-L achieves 87.0% DICE score on BTCV (+1.04% over MedNeXt-L) with 96.5% fewer parameters and 77.03% lower FLOPs. In addition, EfficientMedNeXt-S offers comparable DICE score, improved HD95, and 78.1% higher throughput while reducing parameters by 98.5% and FLOPs by 95%. These results demonstrate EfficientMedNeXt’s efficiency and accuracy, making it well-suited for real-world clinical applications. Code will be released upon acceptance.
1.
Rahman, Md Mostafijur; Marculescu, Radu
EffiDec3D: An Optimized Decoder for High-Performance and Efficient 3D Medical Image Segmentation Conference
Proceedings of the IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR), 2025.
Abstract | Links | BibTeX | Tags: 3D Segmentation, Deep Learning Architecture, Efficient AI, Efficient Decoder, Featured, Medical Image Segmentation
@conference{effidec3d@rahman,
title = {EffiDec3D: An Optimized Decoder for High-Performance and Efficient 3D Medical Image Segmentation},
author = {Md Mostafijur Rahman and Radu Marculescu},
url = {https://openaccess.thecvf.com/content/CVPR2025/html/Rahman_EffiDec3D_An_Optimized_Decoder_for_High-Performance_and_Efficient_3D_Medical_CVPR_2025_paper.html},
year = {2025},
date = {2025-06-10},
urldate = {2025-06-10},
publisher = {Proceedings of the IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR)},
abstract = {Recent 3D deep networks such as SwinUNETR, SwinUNETRv2, and 3D UX-Net have shown promising performance by leveraging self-attention and large-kernel convolutions to capture the volumetric context. However, their substantial computational requirements limit their use in real-time and resource-constrained environments. The high #FLOPs and #Params in these networks stem largely from complex decoder designs with high-resolution layers and
excessive channel counts. In this paper, we propose EffiDec3D, an optimized 3D decoder that employs a channel reduction strategy across all decoder stages, which sets the number of channels to the minimum needed for accurate feature representation. Additionally, EffiDec3D removes the high-resolution layers when their contribution
to segmentation quality is minimal. Our optimized EffiDec3D decoder achieves a 96.4% reduction in #Params and a 93.0% reduction in #FLOPs compared to the decoder of original 3D UX-Net. Similarly, for SwinUNETR and SwinUNETRv2 (which share an identical decoder), we observe reductions of 94.9% in #Params and 86.2% in #FLOPs. Our extensive experiments on 12 different medical imaging tasks confirm that EffiDec3D not only significantly reduces the computational demands, but also maintains a performance level comparable to original models, thus establishing a new standard for efficient 3D medical image segmentation. Our implementation is available at https://github.com/SLDGroup/EffiDec3D.},
keywords = {3D Segmentation, Deep Learning Architecture, Efficient AI, Efficient Decoder, Featured, Medical Image Segmentation},
pubstate = {published},
tppubtype = {conference}
}
Recent 3D deep networks such as SwinUNETR, SwinUNETRv2, and 3D UX-Net have shown promising performance by leveraging self-attention and large-kernel convolutions to capture the volumetric context. However, their substantial computational requirements limit their use in real-time and resource-constrained environments. The high #FLOPs and #Params in these networks stem largely from complex decoder designs with high-resolution layers and
excessive channel counts. In this paper, we propose EffiDec3D, an optimized 3D decoder that employs a channel reduction strategy across all decoder stages, which sets the number of channels to the minimum needed for accurate feature representation. Additionally, EffiDec3D removes the high-resolution layers when their contribution
to segmentation quality is minimal. Our optimized EffiDec3D decoder achieves a 96.4% reduction in #Params and a 93.0% reduction in #FLOPs compared to the decoder of original 3D UX-Net. Similarly, for SwinUNETR and SwinUNETRv2 (which share an identical decoder), we observe reductions of 94.9% in #Params and 86.2% in #FLOPs. Our extensive experiments on 12 different medical imaging tasks confirm that EffiDec3D not only significantly reduces the computational demands, but also maintains a performance level comparable to original models, thus establishing a new standard for efficient 3D medical image segmentation. Our implementation is available at https://github.com/SLDGroup/EffiDec3D.
excessive channel counts. In this paper, we propose EffiDec3D, an optimized 3D decoder that employs a channel reduction strategy across all decoder stages, which sets the number of channels to the minimum needed for accurate feature representation. Additionally, EffiDec3D removes the high-resolution layers when their contribution
to segmentation quality is minimal. Our optimized EffiDec3D decoder achieves a 96.4% reduction in #Params and a 93.0% reduction in #FLOPs compared to the decoder of original 3D UX-Net. Similarly, for SwinUNETR and SwinUNETRv2 (which share an identical decoder), we observe reductions of 94.9% in #Params and 86.2% in #FLOPs. Our extensive experiments on 12 different medical imaging tasks confirm that EffiDec3D not only significantly reduces the computational demands, but also maintains a performance level comparable to original models, thus establishing a new standard for efficient 3D medical image segmentation. Our implementation is available at https://github.com/SLDGroup/EffiDec3D.
