Comparison of denoising tools for the reconstruction of nonlinear multimodal images

Biophotonic multimodal imaging techniques provide deep insights into biological samples such as cells or tissues. However, the measurement time increases dramatically when high-resolution multimodal images (MM) are required. To address this challenge, mathematical methods can be used to shorten the acquisition time for such high-quality images. In this research, we compared standard methods, e.g., the median filter method and the phase retrieval method via the Gerchberg-Saxton algorithm with artificial intelligence (AI) based methods using MM images of head and neck tissues. The AI methods include two approaches: the first one is a transfer learning-based technique that uses the pre-trained network DnCNN. The second approach is the training of networks using augmented head and neck MM images. In this manner, we compared the Noise2Noise network, the MIRNet network, and our deep learning network namely incSRCNN, which is derived from the super-resolution convolutional neural network and inspired by the inception network. These methods reconstruct improved images using measured low-quality (LQ) images, which were measured in approximately 2 seconds. The evaluation was performed on artificial LQ images generated by degrading high-quality (HQ) images measured in 8 seconds using Poisson noise. The results showed the potential of using deep learning on these multimodal images to improve the data quality and reduce the acquisition time. Our proposed network has the advantage of having a simple architecture compared with similar-performing but highly parametrized networks DnCNN, MIRNet, and Noise2Noise.

Automated summary

Biophotonic multimodal imaging techniques are used to study biological samples, but the measurement time increases significantly for high-resolution images. Mathematical methods and artificial intelligence (AI) based approaches were compared in this research to shorten the acquisition time and improve data quality. The AI methods included transfer learning using the pre-trained network DnCNN and training networks using augmented images. Deep learning techniques showed potential in improving image quality and reducing acquisition time. The proposed network had a simpler architecture compared to similar-performing but highly parametrized networks.