Neural Architecture Search for biomedical image classification: A comparative study across data modalities

Deep neural networks have significantly advanced medical image classification across various modalities and tasks. However, manually designing these networks is often time-consuming and suboptimal. Neural Architecture Search (NAS) automates this process, potentially finding more efficient and effective models. This study provides a comprehensive comparative analysis of our two NAS methods, PBC-NAS and BioNAS, across multiple biomedical image classification tasks using the MedMNIST dataset. Our experiments evaluate these methods based on classification performance (Accuracy (ACC) and Area Under the Curve (AUC)) and computational complexity (Floating Point Operation Counts). Results demonstrate that BioNAS models slightly outperform PBC-NAS models in accuracy, with BioNAS-2 achieving the highest average accuracy of 0.848. However, PBC-NAS models exhibit superior computational efficiency, with PBC-NAS-2 achieving the lowest average FLOPs of 0.82 GB. Both methods outperform state-of-the-art architectures like ResNet-18 and ResNet50 and AutoML frameworks such as auto-sklearn, AutoKeras, and Google AutoML. Additionally, PBC-NAS and BioNAS outperform other NAS studies in average ACC results (except MSTF-NAS), and show highly competitive results in average AUC. We conduct extensive ablation studies to investigate the impact of architectural parameters, the effectiveness of fine-tuning, search space efficiency, and the discriminative performance of generated architectures. These studies reveal that larger filter sizes and specific numbers of stacks or modules enhance performance. Fine-tuning existing architectures can achieve nearly optimal results without separating NAS for each dataset. Furthermore, we analyze search space efficiency, uncovering patterns infrequently selected operations and architectural choices. This study highlights the strengths and efficiencies of PBCNAS and BioNAS, providing valuable insights and guidance for future research and practical applications in biomedical image classification.