Abordagem de aprendizado profundo para extração de quadros significativos em volumes de tomografia computadorizada

Abstract

A common approach for analyzing medical images on volumetric data employs deep 2D convolutional neural networks (2D CNN), which imply the use of individual frames. This is largely attributed to the challenges posed by the nature of three-dimensional data: variable volume size, sufficient GPU and RAM allocation, parameter optimization, and more. However, handling the individual frames independently in 2D CNNs deliberately discards the temporal information that constitutes the depth of the volumes, which results in poor performance for the intended task. Therefore, it is important to develop methods that go beyond the computational requirements imposed in order to take advantage of 3D information. For this, we propose an unsupervised method based on Grad-Cam to select key-frames in computed tomography volumes by evaluating the activation map in the last convolutional layer of a CNN3D designed for this purpose. The diagnosis of coronavirus disease was used as a case study for this first stage of the project. Even though the diagnosis by Machine Learning methods already shows promising results through the use of neural networks for the evaluation of radiological images of the lungs, the vast majority of methods use images pre-selected by human professionals to compose an adequate database and this is aggravated when computed tomography volumes are used, where it is necessary to separate more significant frames for clinical evaluation because the full volume analysis is computationally expensive and time-consuming. Extensive experiments with computed tomography volumes demonstrated the success of the proposed methodology. The effectiveness of the Grad-Cam Slice Selection (GSS) method was shown to outperform current state-of-the-art techniques, both in terms of area under the ROC curve (AUC) and F1 Score, in all configurations tested.