Speaker: Ms. Isha
Title: Deep Learning based correction of imaging artifacts in optoacoustic and computed tomography
S. R. Number: 01-02-00-10-12-19-2-17774
Date/Time: March 20th 2025, Thursday / 11:00 AM
Venue: SV Narsaiah Auditorium, IAP Department
Abstract:
Medical imaging techniques such as optoacoustic tomography (OAT), photoacoustic imaging (PAI), and computed tomography (CT) have been instrumental in advancing diagnostic and therapeutic applications. However, these modalities suffer from limitations such as image degradation due to transducer spatial impulse response (SIR), incomplete and noisy data acquisition, and high radiation exposure. This work presents a series of deep learning-based methods to address these challenges and enhance image reconstruction quality across these imaging modalities.
In the first part of this study, we address the degradation of OAT images caused by the spatial impulse response (SIR) of transducer arrays. The SIR effect leads to distorted optoacoustic signals, which, in turn, result in inaccurate pressure distribution estimation and degraded image quality. We modelled and corrected the SIR for two transducer configurations: focused-concave and unfocused-linear arrays. The comparison revealed that the SIR effect was more pronounced in the focused-concave array case. Sparse recovery-based deconvolution and deep learning-based FD-UNet architectures were employed to mitigate these distortions. The FD-UNet model demonstrated a significant improvement in image quality, achieving a 50-80% enhancement in the Structural Similarity Index Measure (SSIM) and more than a 10 dB increase in Peak Signal-to-Noise Ratio (PSNR) compared to traditional backprojection and sparse recovery techniques. Experimental validation further confirmed improved localization accuracy and contrast enhancement of approximately 25%.
The second part of this study focuses on overcoming the challenges associated with photoacoustic imaging (PAI), where reconstruction artifacts arise due to limited transducer counts, incomplete angular coverage, and noise conditions. To tackle this, we introduce a transformer-based dual SwinUNet architecture that effectively learns feature representations from both the image and sinogram domains. The proposed model employs multiple loss functions, including noise-to-signal ratio (NSR) and mean square error (MSE), to optimize the learning process. A contrastive loss function in the sinogram domain is further incorporated to enhance feature discrimination between positive and negative data pairs. Evaluations were conducted under varying conditions, including different transducer counts, angular coverage, and noise levels. The results demonstrated that the proposed network outperformed conventional architectures such as ResNet, UNet, FD-UNet, and TNet. Specifically, when tested on data acquired with 100 transducers covering a 135° angular range, the proposed model improved the universal image quality index (UIQI) by 18% compared to FD-UNet, highlighting its effectiveness in enhancing PAI image reconstruction.
Finally, we address the challenge of low-dose CT (LDCT) imaging, where reducing radiation exposure often results in lower image quality due to increased noise and loss of structural information. To counteract this issue, we propose a multi-encoder single-decoder UNet framework designed to leverage multiple reconstruction algorithms, including backprojection, l2-norm loss, l1-norm loss, and Cauchy loss. Each of these reconstruction algorithms captures distinct features, and their combined integration enhances the quality of the reconstructed images. The proposed framework demonstrated a 10% improvement in PSNR over conventional single-encoder models, confirming its effectiveness in improving LDCT imaging. Furthermore, the model was extended across various deep learning architectures, including ResNet, ResUNet, UNet, and DenoMamba, and was shown to outperform the state-of-the-art DuDoNet network for LDCT. The robustness of the framework was validated using outputs from different scanners and manufacturers. Additionally, an investigation into model complexity revealed that network performance stagnates beyond a specific parameter threshold, highlighting the importance of optimal model design to balance performance and computational efficiency.
In summary, this thesis introduces the potential of deep learning-driven frameworks in mitigating imaging artifacts, enhancing reconstruction accuracy, and optimizing medical imaging workflows. By integrating domain-specific priors with advanced neural architectures, these techniques pave the way for improved diagnostic imaging with reduced data acquisition constraints and lower radiation exposure. The findings of this study contribute to advancing deep learning methodologies for medical imaging applications and open avenues for future research in real-time image reconstruction and clinical translation.