Kernkonzepte
The proposed DensePANet model employs a novel FD-UNet++ architecture in its generator to significantly improve the reconstruction performance of photoacoustic tomography images from sparse data.
Zusammenfassung
The paper presents a new model called DensePANet for accurate photoacoustic tomography (PAT) image reconstruction from sparse data. The key highlights are:
The proposed model uses a novel modification of UNet, called FD-UNet++, as the generator. This architecture combines features from dense blocks and UNet++ to considerably improve the reconstruction performance.
DensePANet is a supervised GAN-based post-processing reconstruction algorithm that leverages the strengths of generative adversarial networks in generating realistic images.
Extensive experiments were conducted on three datasets - Simulated Vessels, Mouse-Abdomen, and Brain Tumor MRI. Quantitative results show that DensePANet outperforms other prevalent deep learning techniques like UNet, UNet++, and FD-UNet in terms of SSIM and PSNR.
The nested skip connections in FD-UNet++ and the integration of dense blocks help capture low-level and high-level features effectively, leading to better artifact removal and higher quality reconstructions.
Compared to other models, DensePANet exhibits lower computational complexity while achieving superior reconstruction performance, making it a promising solution for real-time applications.
Statistiken
The initial pressure distribution was based on publicly available fundus oculi vessel data from the DRIVE database.
The Whole-body mouse dataset had 274 images, which were augmented to 5000 images.
The Brain Tumor MRI dataset had 1321 images, which were divided into 1057 training and 264 testing images.
Zitate
"DensePANet is capable of reducing artifacts of PAT, initiating from under-sampling and the limited-view problem. This method has the ability to significantly improve the quality of the image without making any changes to the imaging system or decreasing the speed of imaging as it is important for real-time imaging applications."