BioNAS: Neural Architecture Search for Multiple Data Modalities in Biomedical Image Classification

Abstract

Neural Architecture Search (NAS) for biomedical image classification has the potential to design highly efficient and accurate networks automatically for tasks from different modalities. This paper presents BioNAS, a new NAS approach designed for multi-modal biomedical image classification. Unlike other methods, BioNAS dynamically adjusts the number of stacks, modules, and feature maps in the network to improve both performance and complexity. The proposed approach utilizes an opposition-based differential evolution optimization technique to identify the optimal network structure. We have compared our methods on two public multi-class classification datasets with different data modalities: DermaMNIST and OrganCMNIST. BioNAS outperforms hand-designed networks, automatic machine learning frameworks, and most NAS studies in terms of accuracy (ACC) and area under the curve (AUC) on the OrganCMNIST and DermaMNIST datasets. The proposed networks significantly outperform all other methods on the DermaMNIST dataset, achieving accuracy improvements of up to 4.4 points and AUC improvements of up to 2.6 points, and also surpass other studies by up to 5.4 points in accuracy and 0.6 points in AUC on OrganCMNIST. Moreover, the proposed networks have fewer parameters than hand-designed architectures like ResNet-18 and ResNet-50. The results indicate that BioNAS has the potential to be an effective alternative to hand-designed networks and automatic frameworks, offering a competitive solution in the classification of biomedical images.