Denoising and enhancement in medical imaging modalities using deep learning

Medical imaging is a crucial component in the modern healthcare system, providing essential information for diagnosing, monitoring, and treating medical conditions. The trade-off between image quality and acquisition time is the key challenge in this area. High-quality imaging often requires prolonged scanning, leading to delays in diagnosis and treatment. State-of-the-art approach to mitigating this challenge is enhancing fast-scanned images via computational methods, which can improve efficiency without compromising quality. In this thesis, the issue is addressed by leveraging deep learning techniques to develop innovative methods for denoising and reconstructing low-quality biological signals and images, focusing on Raman spectroscopy and photoacoustic microscopy (PAM). Raman spectroscopy, while providing rapid and non-invasive molecular information, suffers from weak signal intensity and lengthy exposure times. We propose a fully convolutional encoder-decoder architecture for noise reduction in Raman spectra acquired with tenfold lower exposure times. Our results show superior performance compared to conventional denoising techniques, improving the signal-to-noise ratio by 20% to 80%. PAM combines optical and acoustic imaging for enhanced penetration depth but faces challenges in scanning at high spatial resolution. We introduce DiffPam, a novel algorithm based on diffusion models, which accelerates the PAM imaging without requiring large datasets for training. DiffPam achieves a fivefold increase in scanning speed with minimal information loss, outperforming traditional methods in peak signal-to-noise ratio and structural similarity index. These advancements in Raman spectroscopy and PAM highlight the transformative potential of deep learning to enhance medical imaging quality and efficiency, paving the way for more accurate diagnostics and improved patient care.