In November 2017 a Stanford team led by Pranav Rajpurkar, Jeremy Irvin, and Andrew Ng posted “CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning.” Chest radiographs are among the most common imaging tests in the world, and pneumonia is a frequent and dangerous finding, so automating its detection drew wide attention.
CheXNet is a 121-layer convolutional neural network (a DenseNet) trained on ChestX-ray14, at the time the largest public chest X-ray dataset, with over 100,000 frontal images labeled for 14 thoracic conditions. The model outputs a probability for each disease along with a heatmap highlighting the regions most influencing its decision.
To benchmark it, the authors had four practicing academic radiologists annotate a test set and compared the model’s pneumonia detection against theirs. They reported that CheXNet exceeded the average radiologist’s performance on the F1 metric, a combined measure of precision and recall, and extended the model to achieve strong results across all 14 conditions.
CheXNet became one of the most cited early medical-imaging deep-learning papers and a standard teaching example. It also drew careful scrutiny: the “radiologist-level” claim rested on a single dataset with noisy machine-mined labels and a limited comparison, illustrating early on how benchmark wins in clinical AI need cautious interpretation before they translate to the clinic.