EM-NAS: A Cell-Based Evolutionary Multi-scale Neural Architecture Search for Medical Image Segmentation

Accurate segmentation of medical images is a crucial part of pathology research and clinical practice. Currently, deep neural network (DNN)-based models are commonly used for medicad image segmentation. However, most DNN-based models are manually designed by experts which is a time-consuming and labor-intensive process as well as poorly generalized for different applications. Therefore, we propose EM-NAS, a novel evolutionary multi-scale NAS method with a cell-based design for medical image segmentation. Specifically, the structures of encoder and decoder cells are represented by directed acyclic graphs (DAGs). Moreover, some non-subsampling multi-scale decomposition operations are introduced into the candidate operation set to capture multi-scale, multi-directional and multi-frequency features of medical images. Subsequently, we specify the cell of the network through a well-designed discrete coding strategy, and a GA is used to search for the optimal network architecture. The experimental results demonstrate the effectiveness of the proposed model on two public datasets.