Cross-Domain Transfer Learning for Medical Condition Classification from Infant X-ray Images

Field of artificial intelligence technology has flourished in recent years which led to a creation of diagnosing or assessing diseases using X-ray images. On the downside, these creations focus mostly on adult X-ray images and can not accurately diagnose infant X-ray images. This issue stems from a combination of factors: limited availability of datasets containing infant X-ray and significant variation in these images due to the rapid development of the infant body. Therefore, there is a high demand to develop comprehensive solutions that address these challenges and provide accurate insights. The proposed representation learning-based framework comprises two stages: auto-encoder-based representation learning and transfer learning for diagnosis. The first stage uses adult X-ray images to train the model for improved representation, generating identical reconstructed images. The second stage utilizes pre-trained models to diagnose diseases and predict infant age, enhancing accuracy by accounting for age-related variations in X-ray shapes. This innovative approach represents the first endeavor in unrestricted pediatric X-ray diagnosis, utilizing self-supervised learning for enhanced accuracy. As a result, the comprehensive and extensive experiment allows the proposed method to outperform in comparison to the existing methods. I expect that my research will contribute to the pediatric field of medicine and serve as the foundation of diverging the utility of artificial intelligence.