CT-SCAN

Instructions:

You can upload a chest ct scan image to be investigated by AI for COVID-19 detection.
Accepted image format is jpg, jpeg or png.
The AI will return a result 0 for COVID-19_negative or 1 for COVID-19_positive with a confidence percentile.
The current accuracy percentile for the AI is 91% so there is a possibility for 1 false out of 10 results depending on the ct-scan image quality and clarity. (All results can be correct)

Kindly select an Image

Disclaimer:

COVID-19 Ct-Scan detector is in experiment phase to help fast detection of COVID-19 but the output result can't be considered yet as final confirmed result. Medical reference is the final trusted reference for result confirmation.
You are welcome to challenge the AI with other than ct scan images. but the expected result will probably be always 1, as an unrecognized image were the model is train and recognize ct scan images only.
We appreciate your support and feedback of confirming the result to help us develop the AI.

Summary

First we prepare the data following below steps:

Import COVID-19/Non_COVID-19 images files names into a train DataFrame with their labels. Then coping the last rows for testing DataFrame and delete them from the train df.
Shuffle training dfs, converting each to tensors and join all three train dfs into one big train df shape (3456, 256, 256, 3), and test (700)
Apply Data Augmentation (random rotation, shift, flip, zoom, shear transforms) on all three datasets images generating new 2000 images for generalizing the model as much as possible.
Join new images' dataset, after importing them to df and convert to tensors. the final training/testing dfs (5535, 256, 256, 3) ,and testing (1100, 256, 256, 3).
Import pretrained model “EfficientNetB6” from tensorflow.keras.applications and add our input shape (256,256,3), full connected layers (2048, 512,128) with ReLu as activation function, plus BatchNormalization with dropout (0.1 or 0.2) methods each. Finally, output layer for 2 classes with softmax as activation functions.
Train the data and save the weights for best max val_accurecy or min loss metrics.
Fine tune the model with extra very low learning rate for the optimization function.
Test the result and achieve high accuracy 91% on the testing 1100 images.