Large language models driven neural architecture search for universal and lightweight disease diagnosis on histopathology slide images.

Abstract

Artificial Intelligence has revolutionized healthcare by offering smart services and reducing diagnostic burden, particularly facilitating the identification and segmentation of malignant tissues. However, current task-specific approaches require disease-specific models, while universal foundation models demand costly customization for complex cases, hindering practical deployment in clinical environments. We present Pathology-NAS, a universal and lightweight medical analysis framework that leverages LLMs' knowledge to refine the architecture space across diverse scenarios, eliminating the need for exhaustive search. Pathology-NAS is pretrained on 1.3 million images across three supernet architectures, providing a robust visual foundation that generalizes across diverse tasks. Across breast cancer and diabetic retinopathy diagnosis tasks, Pathology-NAS achieves 99.98% classification accuracy while reducing FLOPs by 45% compared to leading methods. Our model delivers near-optimal architectures in just 10 iterations, bypassing the exponential search space. Pathology-NAS provides accurate tumor recognition across diverse tissues with computational efficiency, making AI-assisted diagnosis practical even in resource-constrained clinical environments.