predator-1-ml / covid

Covid Classification and Opacification using Transfer Learning

The application of computer vision technology to detect and classify this virus from a chest X-ray picture can be a very valuable addition to the less sensitive traditional method of detecting COVID-19, namely reverse transcription polymerase chain reaction (RT-PCR). This automated technique has the potential to improve on current COVID-19 treatment methods while also alleviating the scarcity of skilled physicians in rural areas. Again, segmenting diseased regions from a chest X-ray picture can aid medical practitioners in gaining insight into the afflicted area. So, in this research, we employed a deep learning-based transfer learning approach for CT-scan and X-ray image classification, and a U-Net architecture for segmentation to segment the afflicted region. On the available X-ray dataset, 99.7% classification accuracy was obtained, 99.4% on the available CT-scan dataset, and 87% average accuracy from the segmentation process.

Datasets

We have utilized different datasets for classification and segmentation. For classification, we have used both Covid and Non-Covid pictures of X-ray and CT-scans. We used 13816 images for X-ray, consisting of 3616 Covid-19 and 10200 Non-Covid images. We also used 14500 images for CT-scans, consisting of 7593 Covid-19 and 6893 Non-Covid images.For segmentation, we have used 2 datasets.Both the datasets consists of CT-scans along with their infection masks. In the first dataset, it consists of 1200 images with the infection mask. In the second dataset, it consists of 1564 CT-scans along with their infection masks.

Datasets	Covid	Non-Covid
CT-scans	7593	6893
X-Rays	3616	10200

Datasets	Mask Images
I	1200
II	1564

Methodology

The overall methodology is given in Fig. 1. For classification and segmentation, we used two different approaches. We employed an uncertainity aware transfer learning approach for classification. We utilized the U-net framework for segmentation.

Classification

To detect the presence of COVID-19, the suggested framework only relies on the information content of X-rays and CT images. In this report, we evaluate two best pretrained networks on the ImageNet data set and import and adjust them for COVID-19 identification. ResNet50 and InceptionResNetV2 are the names of these networks.

Segmentation

The general approach to semantically segmenting pictures is to develop a structure that collects features through consecutive convolutions and outputs a segmentation map. U-NET was created with the intention of comprehending and segmenting medical images. It is a significant architecture in the medical imaging automation society and has a wide range of applications in the sector. In this part, we go through the network’s core technological aspects and how they contribute to successful outcomes.

Results

I have done rigourous research on the results part, I have compared my model performance with the other state of the art models. My model outperformed all the models that I have considered on the above mentioned datasets. For more info on results, you can check the report included in the repo.