New patch-based strategy for COVID-19 automatic identification using chest x-ray images

Abstract

Purpose: The development of a robust model for automatic identification of COVID-19 based on chest x-rays has been a widely addressed topic over the last couple of years; however, the scarcity of good quality images sets, and their limited size, have proven to be an important obstacle to obtain reliable models. In fact, models proposed so far have suffered from over-fitting erroneous features instead of learning lung features, a phenomenon known as shortcut learning. In this research, a new image classification methodology is proposed that attempts to mitigate this problem.

Methods: To this end, annotation by expert radiologists of a set of images was performed. The lung region was then segmented and a new classification strategy based on a patch partitioning that improves the resolution of the convolution neural network is proposed. In addition, a set of native images, used as an external evaluation set, is released.

Results: The best results were obtained for the 6-patch splitting variant with 0.887 accuracy, 0.85 recall and 0.848 F1score on the external validation set.

Conclusion: The results show that the proposed new strategy maintains similar values between internal and external validation, which gives our model generalization power, making it available for use in hospital settings.

Supplementary information: The online version contains supplementary material available at 10.1007/s12553-022-00704-4.

Keywords: Automatic classification; COVID-19; Chest X-Rays.

Fig. 1

Division into six regions of a CXR based on lung fields, which is the starting point for manual annotation by radiologists

Fig. 2

Segmentation process and division into patches

Fig. 3

Phase A of the system, division into patches of the images annotated by the radiologists and obtaining Initial Models from the new partitions generated

Fig. 4

Scheme followed to increase the set of annotated images

Fig. 5

Confusion matrix for the different variants of patch splitting in Phase A

Fig. 6

Macrometrics for the different variants of the Phase A patch division

Fig. 7

Confusion matrix for the different patch splitting variants in Phase A for the external evaluation set

Fig. 8

Confusion matrix for the different Phase C patching variants for 10% of the internal validation data

Fig. 9

Macrometrics for the different Phase C patch splitting variants using 10% of the images as evaluation

Fig. 10

Confusion matrix for the different Phase C patching variants for the external evaluation set