Encoding-decoding CNNs for noise reduction to design more interpretable CNNs

For his PhD research, Luis Zavala Mondragon presented an extensive study of encoding-decoding CNNs for noise reduction and employed the findings to design new CNNs that are more interpretable. He presents an approach to understand CNNs that can be applied to design new models where the internal operation is more interpretable and can be optimized. The work analyzes the reconstruction characteristics of CNN models, while considering the implications of their linear and non-linear parts. It is shown that such analyses reveal internal conditions that the models should comply with in order to preserve the signal. This dissertation is among the first in exploring the signal reconstruction behavior of trained ED CNN models and confirm it by a theoretical analysis.

Deep neural networks (DNNs) have faced an exponential increase in data analysis applications and modelling performance. Among other factors, the success of these models is driven by the large availability of data and computing power. The performance growth of these models as well as new striking applications have enabled that DNNs and deep learning (DL) go beyond academic and industrial environments and have now a widespread adoption by the general public in specific daily life applications. Besides the mainstream DL applications provided by text and image generation, DNNs have been applied to other fields such surveillance, autonomous driving and (most relevant for this thesis) medical imaging. With the advent of deep learning, conventional signal/image processing algorithms have been often outperformed and replaced by encoding-decoding (ED) CNNs in tasks such as image denoising.

New developments and theoretical insights

As with all deep learning applications, significant research efforts have been made to enhance the architectures of CNNs. In many cases, these innovations are based on heuristics that offer restricted understanding of the internal operation of these models. In critical applications, such as computed tomography (CT) imaging, it is very important to use well-understood and reliable systems. These requirements have sparked new developments and theoretical insights that explain CNNs from a signal processing perspective, which opens up new exciting opportunities for understanding and improving CNNs.

Lossless signal representations

In order to further advance the understanding in the signal processing behavior of ED CNNs, Zavala Mondragon first revisits signal processing concepts such and the theory of deep convolutional framelets. Based on these elements, he analyzes the encoder-decoder architecture and extends the theory of deep convolutional framelets, by complementing it with concepts such as statistical estimators and Wiener filtering. The explored concepts are used to understand and analyze the signal propagation of CNNs. These analyses result in approaches to modify CNNs such that they become suited for (almost) lossless signal representations, while the non-linear part of the model suppresses the noise. These findings form the basis for modelling contributions in the research.

More efficient and interpretable

Based on these insights noise reduction models were developed which combine elements of wavelet-based denoising algorithms and integrates them with convolutional neural networks to build CNNs which are more efficient and interpretable than conventional models. Finally, in order to show an additional application that leverages the developed models, the last part of the research incorporates elements of the CNNs and integrates them into a model-based dual-energy CBCT material decomposition algorithm.

Title of PhD thesis: Fitting signal processing into CNNs with applications to CT denoising. Supervisors: Prof. Peter de With and Dr. Fons van der Sommen.