3 ms·
As the authors rightly call out in the abstract obtaining large amounts of annotated data poses a challenge for training deep learning models for this purpose B
by yellow_postit 7y ago
As the authors rightly call out in the abstract
obtaining large amounts of annotated data poses a challenge for training deep learning models for this purpose
But doesn't appear that the data they've collected and annotated is made available from my read of the paper, I get that this is from a company (DeepHealth) but it seems like an opportunity for NIH to push for more broadly available data sets.
Anyone have a good reference point for the reader selection of 5 specialists with 5.6yrs avg experience? That population seems small. Another opportunity for licensing bodies or national institutions to grow a publicly available dataset -- including annotations from a wider selection of imaging specialists.
- melan13 7y agoMammogram Images are far away to present any privacy challenge especially if identities are not disclosed.
- 1e-9 7y agoThe radiologists in this study had read 6,969 mammograms on average over the preceding year. That's about 15X the certification requirement and 4X the average for U.S. doctors. Reading volume is one of the main predictors of performance[1], which suggests that these doctors were probably above-average readers. It would have been nice to see more readers involved, but reader studies are a major effort. Even with the small sample size of readers, these results were statistically significant as well as clinically significant. [1] https://www.ncbi.nlm.nih.gov/pubmed/21343539 https://www.ncbi.nlm.nih.gov/pubmed/21343539