An evolutionary block based network for medical image denoising using Differential Evolution

Image denoising is the key component in several computer vision and image processing operations due to unavoidable noise in the image generation process. For medical image processing, deep convolutional neural networks (CNN) gives a state-of-the-art performance. However, network structures are manually constructed for specific tasks and require several trials to tune a large number of hyperparameters, which can take a long time to construct a network. Additionally, the fittest hyperparameters which may be suitable for source data properties like noisy features cannot be easily found to target data. The realistic noise is generally mixed, complex, and unpredictable in medical images, which makes it difficult to design an efficient denoising network. We developed a Differential Evolution (DE) based automatic network evolution model in this paper to optimize the network architectures and hyperparameters by exploring the fittest parameters. Furthermore, we adopted a transfer learning technique to accelerate the training process. The proposed evolutionary algorithm is flexible and finds optimistic network architectures using well-known methods including residual and dense blocks. Finally, the proposed model was evaluated on four different medical image datasets. The obtained results at different noise levels show the potentiality of the proposed model named DEvoNet for identifying the optimal parameters to develop a high-performance denoising network structure.(C) 2022 Elsevier B.V. All rights reserved.

Automated summary

The study proposed an automatic network evolution model based on Differential Evolution for optimizing medical image denoising network structures and hyperparameters. By exploring the fittest parameters and accelerating the training process, the model was evaluated on four different medical image datasets.