Useful references

Best to read

• comparison of different models: https://mc.ai/comparing-chexnet-with-several-classic-neural-networks/
• best medical ai research you've never heard of: https://lukeoakdenrayner.wordpress.com/2019/06/01/the-best-medical-ai-research-that-you-probably-havent-heard-of/
• overview of existing approaches: https://www.nature.com/articles/s41598-019-42294-8
• python env: https://realpython.com/effective-python-environment/

Core model

• original dataset & models: Want at al CVPR 2017: http://openaccess.thecvf.com/content_cvpr_2017/papers/Wang_ChestX-ray8_Hospital-Scale_Chest_CVPR_2017_paper.pdf, announcement: https://www.nih.gov/news-events/news-releases/nih-clinical-center-provides-one-largest-publicly-available-chest-x-ray-datasets-scientific-community
• Stanford paper: https://arxiv.org/pdf/1711.05225.pdf
• dense net in keras: https://keras.io/applications/#densenet
• Stanford landing page: https://stanfordmlgroup.github.io/projects/chexnet/

Implementations

• Keras implementation YES: https://github.com/brucechou1983/CheXNet-Keras
• another comparison project: https://github.com/arnoweng/CheXNet
• pytorch comparison project: https://github.com/evakli11/cs541dlfinalproject_chexnet
• reproducing CheXNet in pytorch: https://github.com/jrzech/reproduce-chexnet/blob/master/model.py

Notebooks

• great walkthrough: https://www.kaggle.com/paultimothymooney/predicting-pathologies-in-x-ray-images
• densenet version: https://www.kaggle.com/ashishpatel26/chexnet-radiologist-level-pneumonia-detection
• basic tutorial: https://towardsdatascience.com/deep-learning-with-x-ray-eyes-eae0ac39b85f

Extensions / clarifications

• google, addressing the problems: https://ai.googleblog.com/2019/12/developing-deep-learning-models-for.html
• blogger on problems with the dataset: https://lukeoakdenrayner.wordpress.com/2017/12/18/the-chestxray14-dataset-problems/
• and on medical AI challenges in general: https://lukeoakdenrayner.wordpress.com/2019/05/04/ten-controversial-opinions-about-medical-ai/ or https://lukeoakdenrayner.wordpress.com/
• criticisms wrt COVID: https://arxiv.org/abs/2004.12823 and https://arxiv.org/abs/2004.05405

Class imbalance

• Facebook/Yannis code for long tail data: https://github.com/facebookresearch/classifier-balancing

Papers to explore

• drop an octave: reducing spatial redundancy (more for model size): https://arxiv.org/pdf/1904.05049.pdf

Datasets

• random 5% sample on Kaggle: https://www.kaggle.com/nih-chest-xrays/sample/kernels
• full large dataset: https://www.kaggle.com/nih-chest-xrays/data
• Carey's compilation in airtable: https://airtable.com/tblo7rn1rOAIvymqg/viwD1riGCLqXuaNhU?blocks=hide

Heat maps/regions

• what are these models are actually learning? https://medium.com/@jrzech/what-are-radiological-deep-learning-models-actually-learning-f97a546c5b98
• tuberculosis specifically: https://www.nature.com/articles/s41598-019-42557-4

Reading xrays in general

• more hands on: https://radiologyassistant.nl/chest/chest-x-ray-basic-interpretation
• A whole chapter: https://www.researchgate.net/publication/226521494_Difficulties_in_the_Interpretation_of_Chest_Radiography
• very detailed: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1071655/

Gender differences

• gender prediction from chest xray is possible with 86.6% accuracy: https://lhncbc.nlm.nih.gov/system/files/SPIE_2018_gender_fullpaper_final.pdf
• some conditions are diagnosed differently, e.g tuberculosis in India
• some conditions present differently, e.g in Vietnam : https://www.ncbi.nlm.nih.gov/pubmed/17366010