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

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• Manual design of deep neural networks for medical image classification is time-consuming and suboptimal.
• Neural Architecture Search (NAS) automates network design, potentially yielding more efficient and effective models.

Purpose of the Study:

• To conduct a comparative analysis of two novel NAS methods, PBC-NAS and BioNAS.
• To evaluate their performance on biomedical image classification tasks using the MedMNIST dataset.
• To assess classification accuracy, AUC, and computational complexity (FLOPs).

Main Methods:

• Comparative analysis of PBC-NAS and BioNAS on MedMNIST dataset.
• Evaluation metrics: Accuracy (ACC), Area Under the Curve (AUC), Floating Point Operation Counts (FLOPs).
• Ablation studies on architectural parameters, fine-tuning, search space efficiency, and discriminative performance.

Main Results:

• BioNAS models slightly outperformed PBC-NAS in accuracy (BioNAS-2: 0.848 avg ACC).
• PBC-NAS models demonstrated superior computational efficiency (PBC-NAS-2: 0.82 GB avg FLOPs).
• Both methods surpassed ResNet, AutoML frameworks, and most prior NAS studies in ACC and AUC.

Conclusions:

• PBC-NAS and BioNAS offer efficient and effective automated solutions for biomedical image classification.
• Larger filter sizes and specific module configurations enhance performance.
• Fine-tuning existing architectures without dataset-specific NAS is effective.