Focusing on correct regions: self-supervised pre-training in lung disease classification

Abstract

While the current state of lung disease detection using AI relies heavily on specific data types, limiting its real-world applicability, this work explores leveraging Transfer Learning (TL) with VicReg for improved performance. By using a public Chest X-ray dataset, the proposed model employs a ResNet50 model archi tecture that seamlessly integrates transfer learning and fine-tuned self-supervised Convolutional Neural Networks (CNNs). Can Artificial Intelligence (AI) for lung disease detection be improved to work across different types of medical images? This study addresses this challenge by proposing DOCTOR, a reusable AI model that leverages transfer learning and fine-tuned CNNs. DOCTOR is trained on chest X-rays but designed to be adaptable to other radiology images like CT scans. The results obtained from this proposed model are remarkable, achieving an impressive accuracy rate of 97.37%, sensitivity of 96.30, specificity of 97.30% each, and precision of 96.30%. These exceptional performance metrics demonstrate the proposed model’s exceptional competence and efficacy in accurately detecting lung dis- eases. The trained CNN model utilized a ResNet50 backbone pre-trained using VicReg for robust lung disease detection across various modalities, which is referred to in this paper as DOCTOR.