Medical imaging has been one of the main tools employed during the COVID-19 pandemic for diagnosis and disease progression assessment. The most commonly used have been Chest X-Rays (CXR) and Computed Tomography (CT). However, CXR has a limited sensibility, while CT is more expensive, less accessible, gives more dose to the patients, and requires sanitizing the scanner after each patient acquisition. Tomosynthesis, which obtains X-rays images from a few source positions, has been proposed as a good compromise between both modalities.
The use of Artificial Intelligence (AI) tools to analyze medical images of COVID-19 patients has been proposed by many groups. It has been shown that Neural Networks (NN) can be trained to detect COVID-19 affections accurately provided enough cases are available. Nevertheless, while many public databases of CXR and CT images of COVID-19 patients have been generated worldwide, there is a lack of databases of tomosynthesis images, which makes it difficult to train a NN for this modality.
In this work we propose to use the existing CT and X-ray databases to perform realistic simulations and generate X-Ray tomosynthesis images. We made use of a database containing 200 CT images of COVID-19 patients, along with the segmentations of the lung affected region. Projections at 0⁰ and ±15⁰ were simulated in an in-house developed, GPU-accelerated, ultrafast Monte Carlo (MC) code. Two NN were trained to detect whether each lung is affected by COVID-19 or not: the first one is defined with one input channel corresponding to the 0⁰ projection (which corresponds to a standard CXR), while the other one employs three input channels corresponding to 0⁰ and ±15⁰ projections (which corresponds to a simplified tomosynthesis acquisition). Results show that the three-channel NN outperforms the one-channel NN. Despite the limited number of cases used in this work, and the reduced number of projections, the results are very promising, and motivates further research on the advantages which can be obtained with Tomosynthesis.