Skip to content

Latest commit

 

History

History
 
 

CXRP

Model and Hyper-parameters

The CXRP- xB model - Covid X-ray Profiler with x blocks - was designed with the objective of being specially efficient at analyzing chest x-rays. As such, the architecture is composed of the following:

CXRP Architecture

The number of blocks needed was found through experiments, being 3 blocks the final amount.

Various versions of this architecture were studied, which we show along this document.

Another important aspect of this model is that every execution of the training was made with different weights on the 3 classes, having COVID-19 3x the normal weight of the other two classes. We found this way the model reached better results in less time, learning better the COVID-19 patterns, which was always the worst performing class on all the other models used.

First iteration of the CXRP, CXRP-1B

- 1 Block
- Adam optimizer with learning rate 0.001

Link to access the model weights: https://drive.google.com/drive/folders/1-U4gLUbzta6MbzTPt65MKoLa0OzkRcl4?usp=sharing

Results

Accuracy during Training

Loss during Training

Classification Report

Classification Report

Confusion Matrix

Confusion Matrix

Second iteration of the CXRP, CXRP-2B

- 2 Blocks
- Adam optimizer with learning rate 0.001

Link to access the model weights: https://drive.google.com/drive/folders/1-wAjRyv6ynX3gaWbeKDh-8-pjjJCJlEG?usp=sharing

Results

Accuracy during Training

Loss during Training

Classification Report

Classification Report

Confusion Matrix

Confusion Matrix

Thrid iteration of the CXRP, CXRP-3B

- 3 Blocks
- Adam optimizer with learning rate 0.001

Link to access the model weights: https://drive.google.com/drive/folders/10AmCEvtAPPwVnnVb2-4R44B_hly5IXDk?usp=sharing

Results

Accuracy during Training

Loss during Training

Classification Report

Classification Report

Confusion Matrix

Confusion Matrix

Conclusion

As said earlier, 3 blocks was the best performing setting for this model, as well as the 256-128-3 fully connected layers at the end of the model. Different combinations were tried (more on that on the .ipynb ) but the performance were in some cases slighty worse or even non-usuable on others. This combinations ranged from different kernel sizes, to different versions of the blocks, as well as different fully connected layers on the output of the model. Our objective with this model was to try maximize the recall on the Covid-19 class, without worsening the others, which we thought was well executed.

To test this model, simply run this notebook.