89 — CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning

Using the ~100K-image ChestX-ray14 dataset, this team was able to design a 121-layer CNN that generates a heatmap (probability) of pneumonia, overlaid over a chest X-ray.

The algorithm was compared against a n=420 (niice) training set ground-truthed by a team of four radiologists. The performance of each radiologist was determined by the agreement with the remaining three (to compare the machine learning approach to humans).

The authors trained the CNN on these results from the training set, and found that the system outperformed the humans on a holdout set.

The radiologists, unlike the algorithm, were also asked to annotate all 14 diseased-states found in the dataset; Rajpurkar et al also found that, when trained on these labels, CheXNet also outperformed humans on just about every other disease detection task as well, attaining >70% detection rates on all diseases, and reaching 93% detection rates for Hernia and Emphysema conditions.

The project page is available here, but I couldn’t find links to any available code.