Hey everyone -
I've been working to reproduce [CheXNet](https://arxiv.org/pdf/1711.05225.pdf) - a fantastic paper describing research on a model capable of radiologist-grade pathology classification!
CheXNet uses Class Activation Mappings (CAMs for short) to generate heatmaps that identify what parts of the image the model uses to base its classification. In my case, I'm facing a bit of a struggle reproducing them - most of our classifications are derived from the diaphragm, instead of regions within the lung. Curiously, we are attaining a reasonable AUROC, with .773 on training and .749 on validation data - the paper reports .8062 AUROC.
My current model is being trained on a subsample of the main dataset, and I'm basically looking to this as a way to validate the architecture. I'd love to know if anyone has experienced similar issues and solved them, and could have any input here as well.
If you have a moment to spare - I'd be super grateful for some help from the hackernews community in solving the inaccurate localization issue! [https://dagshub.com/nirbarazida/Pneumonia-Classification/issues/58]
An incorrect localization, despite a correct classification: https://dagshub.com/nirbarazida/Pneumonia-Classification/raw/4c1414d3cd5cf8c693a4f7931843495bd4d96751/evaluating/heatmap_eval/00005532_000_Cardiomegaly.png