Challenges of AI driven diagnosis of chest X-rays transmitted through smart phones: a case study in COVID-19

Abstract Healthcare delivery during the initial days of outbreak of COVID-19 pandemic was badly impacted due to large number of severely infected patients posing an unprecedented global challenge. Although the importance of Chest X-rays (CXRs) in meeting this challenge has now been widely recognized, speedy diagnosis of CXRs remains an outstanding challenge because of fewer Radiologists. The exponential increase in Smart Phone ownership globally, including LMICs, provides an opportunity for exploring AI-driven diagnostic tools when provided with large volumes of CXRs transmitted through Smart Phones. However, the challenges associated with such systems have not been studied to the best of our knowledge. In this paper, we show that the predictions of AI-driven models on CXR images transmitted through Smart Phones via applications, such as WhatsApp, suffer both in terms of Predictability and Explainability, two key aspects of any automated Medical Diagnosis system. We find that several existing Deep learning based models exhibit prediction instability–disagreement between the prediction outcome of the original image and the transmitted image. Concomitantly we find that the explainability of the models deteriorate substantially, prediction on the transmitted CXR is often driven by features present outside the lung region, clearly a manifestation of Spurious Correlations. Our study reveals that there is significant compression of high-resolution CXR images, sometimes as high as 95%, and this could be the reason behind these two problems. Apart from demonstrating these problems, our main contribution is to show that Multi-Task learning (MTL) can serve as an effective bulwark against the aforementioned problems. We show that MTL models exhibit substantially more robustness, 40% over existing baselines. Explainability of such models, when measured by a saliency score dependent on out-of-lung features, also show a 35% improvement. The study is conducted on WaCXR dataset, a curated dataset of 6562 image pairs corresponding to original uncompressed and WhatsApp compressed CXR images. Keeping in mind that there are no previous datasets to study such problems, we open-source this data along with all implementations