COVID-19 detection in CT images with deep learning: A voting-based scheme and cross-datasets analysis

Abstract

Early detection and diagnosis are critical factors to control the COVID-19 spreading. A number of deep learning-based methodologies have been recently proposed for COVID-19 screening in CT scans as a tool to automate and help with the diagnosis. These approaches, however, suffer from at least one of the following problems: (i) they treat each CT scan slice independently and (ii) the methods are trained and tested with sets of images from the same dataset. Treating the slices independently means that the same patient may appear in the training and test sets at the same time which may produce misleading results. It also raises the question of whether the scans from the same patient should be evaluated as a group or not. Moreover, using a single dataset raises concerns about the generalization of the methods. Different datasets tend to present images of varying quality which may come from different types of CT machines reflecting the conditions of the countries and cities from where they come from. In order to address these two problems, in this work, we propose an Efficient Deep Learning Technique for the screening of COVID-19 with a voting-based approach. In this approach, the images from a given patient are classified as group in a voting system. The approach is tested in the two biggest datasets of COVID-19 CT analysis with a patient-based split. A cross dataset study is also presented to assess the robustness of the models in a more realistic scenario in which data comes from different distributions. The cross-dataset analysis has shown that the generalization power of deep learning models is far from acceptable for the task since accuracy drops from 87.68% to 56.16% on the best evaluation scenario. These results highlighted that the methods that aim at COVID-19 detection in CT-images have to improve significantly to be considered as a clinical option and larger and more diverse datasets are needed to evaluate the methods in a realistic scenario.

Keywords: COVID-19; Chest radiography; Deep learning; EfficientNet; Pneumonia.

Fig. 1

Examples of CT images that are: positive for COVID-19 (top) and non-COVID-19 (bottom) from SARS-CoV-2 CT-scan dataset.

Fig. 2

Comparison among different contrast in images.

Fig. 3

Examples of CT images that are: positive for COVID-19 (top) and non-COVID-19 (bottom) from COVID-CT dataset.

Fig. 4

Example of images with textual information.

Fig. 5

MBConv Block [16]. DWConv stands for depthwise conv, k3x3/k5x5 defines the kernel size, BN is batch normalization, $HxWxF$ represents the tensor shape (height, width, depth), and x1/2/3/4 is the multiplier for number of repeated layers.

Fig. 6

Efficient net compound scaling on three parameters (Adapted from Ref. [17]).

Fig. 7

Data augmentation applied using the Augmentor python package. The transformations applied to the images are: rotation (0–15° clockwise or anticlockwise), 20% Zoom or horizontal flipping. All or none changes may be applied/combined according to a probability. (Figure created by the authors).

Fig. 8

ROC curve of the proposed approach.

Fig. 9

Training loss curve of Architecture 2 with input of 500x500.