A stacked ensemble for the detection of COVID-19 with high recall and accuracy

Abstract

The main challenges for the automatic detection of the coronavirus disease (COVID-19) from computed tomography (CT) scans of an individual are: a lack of large datasets, ambiguity in the characteristics of COVID-19 and the detection techniques having low sensitivity (or recall). Hence, developing diagnostic techniques with high recall and automatic feature extraction using the available data are crucial for controlling the spread of COVID-19. This paper proposes a novel stacked ensemble capable of detecting COVID-19 from a patient's chest CT scans with high recall and accuracy. A systematic approach for designing a stacked ensemble from pre-trained computer vision models using transfer learning (TL) is presented. A novel diversity measure that results in the stacked ensemble with high recall and accuracy is proposed. The stacked ensemble proposed in this paper considers four pre-trained computer vision models: the visual geometry group (VGG)-19, residual network (ResNet)-101, densely connected convolutional network (DenseNet)-169 and wide residual network (WideResNet)-50-2. The proposed model was trained and evaluated with three different chest CT scans. As recall is more important than precision, the trade-offs between recall and precision were explored in relevance to COVID-19. The optimal recommended threshold values were found for each dataset.

Keywords: COVID-19; False negatives; Recall; Transfer learning.

Fig. 1

COVID-19 Negative CT scan images.

Fig. 2

COVID-19 Positive CT scan images.

Fig. 3

Part 1 Model architecture.

Fig. 4

Part 2 Model architecture.

Fig. 5

Part 3 Model architecture.

Fig. 6

Combined Model architecture.

Fig. 7

Variation of Precision, Recall, Accuracy and F1 score with threshold on COVID-CT Dataset [88].

Fig. 8

Variation of Precision, Recall, Accuracy and F1 score with threshold on COVID-CTset [89].

Fig. 9

Variation of Precision, Recall, Accuracy and F1 score with threshold on SARS-CoV-2 CT scan dataset [90].

Fig. 10

F1 Score of Different models on different datasets.

Fig. 11

Accuracy of Different models on different datasets.

Fig. 12

Variation of Accuracy with threshold for each dataset.