4S-DT: Self Supervised Super Sample Decomposition for Transfer learning with application to COVID-19 detection

Abstract

Due to the high availability of large-scale annotated image datasets, knowledge transfer from pre-trained models showed outstanding performance in medical image classification. However, building a robust image classification model for datasets with data irregularity or imbalanced classes can be a very challenging task, especially in the medical imaging domain. In this paper, we propose a novel deep convolutional neural network, we called Self Supervised Super Sample Decomposition for transfer learning (4S-DT) model.4S-DTencourages a coarse-to-fine transfer learning from large-scale image recognition tasks to a specific chest X-ray image classification task using a generic self-supervised sample decomposition approach. Our main contribution is a novel self-supervised learning mechanism guided by a super sample decomposition of unlabelled chest X-ray images. 4S-DT helps in improving the robustness of knowledge transformation via a downstream learning strategy with a class-decomposition layer to simplify the local structure of the data.4S-DT can deal with any irregularities in the image dataset by investigating its class boundaries using a downstream class-decomposition mechanism. We used 50,000 unlabelled chest X-ray images to achieve our coarse-to-fine transfer learning with an application to COVID-19 detection, as an exemplar.4S-DThas achieved an accuracy of97.54%in the detection of COVID-19 cases on an extended test set enriched by augmented images, out of which all real COVID-19 cases were detected, which was the highest accuracy obtained when compared to other methods.

A Complete Survey on Automatically Diagnosing COVID-19 In the field of Computer Vision and A Collection of Medical Images

“…Using 8 different Deep Learning Model [4] researchers found NasNet and MobileNet outperformed all the models and they used both CT image and X-ray images to evaluate their system.DenseNets concatenates output features rather than sum the output feature maps of the layer with the incoming feature maps.DenseNet outperformed than other models [18] having 85% sensitivity.In the following section, some other feature extraction methods adopted by the researchers are discussed elaborately. [4], [7], [10], [13], [14], [18], [20], [22], [27], [28], [31], [32], [38], [40], [41], [44], [ 47], [51], [61], [62], [66], [67], [71], [73], [74] 26…”

“…GoogLeNet [7], [27], [40], [44], [66], 5 [49], and a Capsule Network-based Framework (COVID-CAPS) [57].…”

“…Flipping or Rotating [1], [2][15], [16], [21], [22], [24], [25], [27], [35], [36], [37], [38], [42], [43], [4 6], [47], [48], [50], [51], [53], [55], [57], [61], [62], [65], [71], [74] 29…”

“…However, Binary classification may create ambiguity while detecting COVID-19 as it can not distinguish between other Viral Pneumonia and COVID-19. [9], [11,][13], [14], [15], [18] [19], [21], [22], [23], [25], [27], [28], [29], [30], [32], [34] , [35], [37], [42], [44], [46], [ 48], [ 49], [ 51], [52], [ 53], [54], [ 55], [57], [58], [60], [62], [64], [66], [67], [ 68], [69], [72] 40…”

A Comprehensive Survey of COVID-19 Detection Using Medical Images

“…From both Fig. 4 [28] 153 N/A 203 N/A [11] 1126 N/A 938 N/A [12] 6000 N/A 6000 N/A [13] 247 N/A 178 N/A [14] NS NS NS N/A [74] N/A N/A 150 N/A [75] 75541 N/A 64771 N/A [76] 397 N/A 349 N/A [77] 195 N/A 275 N/A [18] 339 COVID-19 images, other lung disease images like Viral pneumonia ( [3], [25], [19], [46], [20]), Bacterial pneumonia ( [60], [19], [61], [62], [20], [63]), fungal pneumonia [64], SARS ( [65], [66], [67], [57]), MERS [66], Influenza [10], Tuberculosis ( [61], [67], [68]) and images of healthy people. The distribution of different types of lung disease or normal images and the number of CT images used by papers are illustrated in Table 2.…”