AshuMaths1729 / COVID-19_XRay_Classifier

As part of the SAMHAR-COVID-19 Hackathon

COVID-19 XRay Classifier

As part of the SAMHAR-COVID-19 Hackathon

The project aims to predict if a person is Normal or having normal Pneumonia or is suffering from COVID-19 Pneumonia.

---

Combined datasets from two sources:

1. Joseph Paul Cohen and Paul Morrison and Lan Dao COVID-19 image data collection, arXiv:2003.11597, 2020 https://github.com/ieee8023/covid-chestxray-dataset https://arxiv.org/abs/2003.11597

2. Kermany, Daniel S., et al. "Identifying medical diagnoses and treatable diseases by image-based deep learning." Cell 172.5 (2018): 1122-1131. https://doi.org/10.1016/j.cell.2018.02.010. https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia

-> Modified data from source 1. Put COVID-19 and COVID-19, ARDS into one group - COVID-19. Put No Finding tagged images under Normal. Put all other images other than these tags, under Pneumonia.

-> Put source 2 images which are under Pneumonia and Normal, under respective categories. Chose very fewer images from source 2, to enable balanced dataset.

-> Will be better if we get more dataset regarding COVID-19 cases, because we do have data regarding Pneumonia and Normal, but nor so much regarding COVID-19.

-> Trained a custom CNN classifier to achieve the task of classfication on three classes of images.

-> The Trained model weights can be found here: https://drive.google.com/file/d/1bvRqYpnxW4QDZoRIWvBdu1sg5ci-yAPR/view?usp=sharing

---

Training Stats

As can be inferred from the plots that accuracy and loss are unusual low and high respectively, which is not acceptable.

This is because we have trained the model on just ~96 training and ~30 validation instances.

Though the data from source 2 did have a lot more images which gives high accuracy when trained, but here since COVID-19 related radiography images which are in source 1, and are 138 in number (PA View Xrays).

So, we decreased the number of images for other categories too, to avoid bias towards other two types of classes, viz. Normal and Pneumonia, rendering our task worthless.

---

To-do:

• Train on more number of instances (~10000) or so, to get better model's performace.
• Use various CNN Architectures to get a better model.
• Deploy the trained classifier on to an Android App, for easy functionality.

---

Team A_Cube

• Ashutosh Agrahari
• Ankur Veer
• Anshuman Singh