TY  - JOUR
AU  - Khan, Romaan
AU  - Uddin, Islam
AU  - Khan, Salman
AU  - Allah, Abdul
PY  - 2026
DA  - 2026/08/11
TI  - Hybrid CNN–Transformer Framework for Structure-Preserving Enhancement of Coronary Angiography Images
JO  - ICCK Transactions on Applied Intelligence and Cybernetics
T2  - ICCK Transactions on Applied Intelligence and Cybernetics
JF  - ICCK Transactions on Applied Intelligence and Cybernetics
VL  - 1
IS  - 2
SP  - 96
EP  - 111
DO  - 10.62762/TAIC.2026.500998
UR  - https://www.icck.org/article/abs/TAIC.2026.500998
KW  - deep learning
KW  - coronary angiography
KW  - image enhancement
KW  - hybrid CNN–Transformer
KW  - structural preservation
KW  - convolutional neural networks
AB  - Coronary artery disease (CAD) remains a leading global cause of mortality, with X-ray coronary angiography serving as the clinical gold standard for evaluating coronary morphology and stenosis. However, angiographic images are often compromised by low contrast, noise, motion blur, and compression artifacts that obscure fine vascular structures and impede both diagnostic interpretation and automated analysis. To address these limitations, this study proposes a hybrid deep learning framework integrating convolutional neural networks (CNNs) with Transformer-based self-attention mechanisms for coronary angiography enhancement. The CNN component targets local feature extraction to suppress noise, enhance edges, and restore vessel details, while the Transformer module captures long-range spatial dependencies to preserve global anatomical continuity and structural coherence. The framework is trained in a supervised manner using paired high-quality and synthetically degraded images from the ARCADE dataset. Quantitative evaluation shows that the proposed hybrid model achieves a Peak Signal-to-Noise Ratio (PSNR) of 36.10 dB and Structural Similarity Index Measure (SSIM) of 0.9670, outperforming the CNN-only baseline (33.69 dB PSNR, 0.9594 SSIM) and conventional methods such as histogram equalization (28.4 dB, 0.87 SSIM) and CLAHE (29.6 dB, 0.90 SSIM). The hybrid approach yields a 2.41~dB PSNR improvement over the CNN baseline, indicating superior reconstruction fidelity and structural preservation. Qualitative assessment further confirms enhanced vessel clarity, contrast, and continuity of thin distal coronary branches. These findings demonstrate the effectiveness of global–local feature fusion for robust and clinically meaningful angiographic image enhancement.
SN  - 3143-0309
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Khan2026Hybrid,
  author = {Romaan Khan and Islam Uddin and Salman Khan and Abdul Allah},
  title = {Hybrid CNN–Transformer Framework for Structure-Preserving Enhancement of Coronary Angiography Images},
  journal = {ICCK Transactions on Applied Intelligence and Cybernetics},
  year = {2026},
  volume = {1},
  number = {2},
  pages = {96-111},
  doi = {10.62762/TAIC.2026.500998},
  url = {https://www.icck.org/article/abs/TAIC.2026.500998},
  abstract = {Coronary artery disease (CAD) remains a leading global cause of mortality, with X-ray coronary angiography serving as the clinical gold standard for evaluating coronary morphology and stenosis. However, angiographic images are often compromised by low contrast, noise, motion blur, and compression artifacts that obscure fine vascular structures and impede both diagnostic interpretation and automated analysis. To address these limitations, this study proposes a hybrid deep learning framework integrating convolutional neural networks (CNNs) with Transformer-based self-attention mechanisms for coronary angiography enhancement. The CNN component targets local feature extraction to suppress noise, enhance edges, and restore vessel details, while the Transformer module captures long-range spatial dependencies to preserve global anatomical continuity and structural coherence. The framework is trained in a supervised manner using paired high-quality and synthetically degraded images from the ARCADE dataset. Quantitative evaluation shows that the proposed hybrid model achieves a Peak Signal-to-Noise Ratio (PSNR) of 36.10 dB and Structural Similarity Index Measure (SSIM) of 0.9670, outperforming the CNN-only baseline (33.69 dB PSNR, 0.9594 SSIM) and conventional methods such as histogram equalization (28.4 dB, 0.87 SSIM) and CLAHE (29.6 dB, 0.90 SSIM). The hybrid approach yields a 2.41~dB PSNR improvement over the CNN baseline, indicating superior reconstruction fidelity and structural preservation. Qualitative assessment further confirms enhanced vessel clarity, contrast, and continuity of thin distal coronary branches. These findings demonstrate the effectiveness of global–local feature fusion for robust and clinically meaningful angiographic image enhancement.},
  keywords = {deep learning, coronary angiography, image enhancement, hybrid CNN–Transformer, structural preservation, convolutional neural networks},
  issn = {3143-0309},
  publisher = {Institute of Central Computation and Knowledge}
}