RespiraAI: Novel Elevated Feature Recalibration Convolutional Neural Network for Efficient Diagnosis and Classification of Pulmonary Diseases

Pulmonary diseases are a group of conditions that prevent the respiratory system from functioning properly. They are the third leading cause of death globally, encompassing a wide range of diseases from viral and bacterial infections to chronic and/or genetic diseases. Traditionally, pulmonary diseases have been diagnosed through chest x-rays/CT scans, and human examination. The error rate of this approach, especially when concerning diseases that present similarly on scans, can be very high. Due to this, several computational models and algorithms have been developed to analyze scans, but deep learning approaches, although they show a lot of promise, can be very computationally expensive and time consuming to train and test. Additionally, current deep learning models have been shown to be insufficient in regard to accounting for heterogeneity among different scans of the same class as well as inaccurate in diagnosing rare pulmonary diseases and differentiating between visually ambiguous scans of different classes due to the imbalance in data coupled with over-focusing on irrelevant portions of the scan. To address these issues, we propose a feature recalibration-based approach to enhance the ability of the model to focus on the most relevant portions of the image, enabling detection of more subtle features, fast training time, and greater accuracy in distinguishing between look alike images of different classes. Thus, we present RespiraAI, a novel CNN architecture integrated into a websitebased platform, for accurate and rapid classification of several types and subtypes of lung disease. RespiraAI displays a training and validation accuracy of 96% and a testing accuracy of 95%, outperforming several state-of-the-art models even though it trained on a small dataset, demonstrating a key feature enabled by its feature recalibration-based approach. Furthermore, because RespiraAI can achieve such a high accuracy with very little data, training takes up just 5-6 GB of RAM, which is over two times less than the average neural network training requirement of 16 GB, and the model is able to classify a single scan almost immediately. The model is also integrated into a website format, making it easy to access for clinics and hospitals all over the world, providing a novel and robust tool for healthcare organizations in underprivileged communities.