New AI Model Developed to Overcome Flaws in Smartphone-Based COVID-19 X-ray Diagnosis

by

Researchers have published a study in Scientific Reports that evaluates the limitations of using chest X-rays (CXR) shared through smartphone applications for COVID-19 diagnosis. While smartphone-based AI models have been developed to automate COVID-19 diagnosis, the study highlights the limitations of these models, particularly when analyzing highly compressed images. To overcome these challenges, the researchers introduce a multi-task learning (MTL) model aimed at accurate COVID-19 diagnosis, even under conditions of image compression.

Before the availability of clinical diagnostic COVID-19 test kits, CXR was commonly used as a triage assessment for the disease. However, with the rapid spread of COVID-19, there was a shortage of radiologists to analyze the CXR images, especially in low-to-middle-income countries and rural areas. To address this issue, AI-based systems were developed to automate COVID-19 diagnosis from CXR images, using smartphones as the medium of implementation. Smartphones offer high-resolution cameras and messaging applications that allow for remote image sharing.

One limitation of using smartphones for AI-aided diagnosis is the loss of image quality due to compression. Although this does not significantly affect diagnoses by expert radiologists, it can significantly alter the performance of AI diagnostic models. The study developed a novel CXR image dataset called ‘WhatsApp CXR’ (WaCXR) to evaluate the effects of compression on AI model performance. The researchers identified two main limitations of current AI systems: ‘Prediction Instability’ (PIP) and ‘Out of Lung Saliency’ (OLS).

PIP refers to the lack of congruency in model predictions between compressed and uncompressed CXR images, which renders the predictions unreliable. OLS is the observation that COVID-19 predictions by AI models are based on CXR image regions outside the lung. To address these limitations, the researchers developed a multi-task learning model called COVIDMT, which showed improved performance compared to current AI models.

The study highlights the need for AI models that can overcome the challenges of compression and ensure reliable and accurate COVID-19 diagnosis. The development of the COVIDMT model represents a step forward in addressing these limitations and improving the performance of AI-based diagnostic systems.

*Note:

  1. Source: Coherent Market Insights, Public sources, Desk research
  2. We have leveraged AI tools to mine information and compile it