Tree-Based Codification in Neural Architecture Search for Medical Image Segmentation
提出一种基于树的遗传编程算法NASGP-Net，以U-Net为骨干自动搜索紧凑的卷积神经网络架构，在多个医学图像分割任务上取得优异结果且参数更少。
Convolutional neural networks (CNNs) have shown a competitive performance in medical imaging applications, such as image segmentation. However, choosing an existing architecture capable of adapting to a specific dataset is challenging and requires design expertise. Neural architecture search (NAS) is employed to overcome these limitations. NAS uses techniques to design the Neural Networks architecture. Typically, the models’ weights optimization is carried out using a continuous loss function, unlike model topology optimization, which is highly influenced by the specific problem. Genetic programming (GP) is an evolutionary algorithm (EA) capable of adapting to the topology optimization problem of CNNs by considering the attributes of its representation. A tree representation can express complex connectivity and apply variation operations. This article presents a tree-based GP algorithm for evolving CNNs based on the well-known U-Net architecture producing compact and flexible models for medical image segmentation across multiple domains. This proposal is called NAS / GP / U-Net (NASGP-Net). NASGP-Net uses a cell-based encoding and U-Net architecture as a backbone to construct CNNs based on a hierarchical arrangement of primitive operations. Our experiments indicate that our approach can produce remarkable segmentation results with fewer parameters regarding fixed architectures. Moreover, NASGP-Net presents competitive results against NAS methods. Finally, we observed notable performance improvements based on several evaluation metrics, including dice similarity coefficient (DSC), intersection over union (IoU), and Hausdorff distance (HD).