Awesome. I love to see small teams tackling big problems like this. Good luck guys!
thank you!!
Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
11–20 of 26 posts
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#12What incentive is there for medical students to do this work for free? I don't know of any doctor that would just give away all their work for learning. They get plenty of that during their lecture.
Is this supposed to be a product to help ai radiology startups curate and manage their data? If so, are we talking about semantic segmentation, localization, or what sort of label? A lot of the time providing explicit data information will require fewer studies to generalize, but require much more work from labeling side.
I would also wonder about your data sourcing. Just because you don't have an FDA product, doesn't mean you're clear of HIPPA rules. Medical images may contain PII, especially scans that include the face.
Edit: couple of typos...
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#13What incentive is there for medical students to do this work for free? I don't know of any doctor that would just give away all their work for learning. They get plenty of that during their lecture. Is this supposed to be a product to help ai radiology startups curate and manage their data? If so, are we talking about semantic segmentation, localization, or what sort of label? A lot of the time providing explicit dat…
Our users participate to win cash prizes, to learn, and compete with others. Everyone loves seeing their name on a leaderboard, especially when they can improve their skills and win cash prizes at the same time!
Today we’re building out our annotation tools. We have bounding boxes, localization, and more depending on the task.
Today we use publicly available datasets and depend on our clients to only provide deidentified images. We’re also working to verify certain users to enable them to see cases with PII.
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#14I would think you could go a long way with just creating an ESP style game: https://en.wikipedia.org/wiki/ESP_game
In this talk, the creator also goes into how they created a version to actually locate the key points in the image: https://www.youtube.com/watch?v=tx082gDwGcM
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#15I would think you could go a long way with just creating an ESP style game: https://en.wikipedia.org/wiki/ESP_game In this talk, the creator also goes into how they created a version to actually locate the key points in the image: https://www.youtube.com/watch?v=tx082gDwGcM
Yes I am a huge fan of his work! He went on to found DuoLingo, which has inspired us a lot. I recommend everyone watch the talk you linked and also his TED
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#16>they could reliably distinguish cancerous moles from benign ones better than individual dermatologists
Do you have anything published about this?
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#17>they could reliably distinguish cancerous moles from benign ones better than individual dermatologists Do you have anything published about this?
Not yet! I’ll have a paper on skin cancer diagnosis out soon though, stay tuned or reach out directly for more info!
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#18Tangentially related, it kind of reminded me of Snapshot Serengeti which labels images with animals. Interestingly, these labels aren't used to train models. They're just used to filter pictures taken by automated motion sensor based cameras. The pictures with actual animals in them are used to study animal migrations.
https://www.zooniverse.org/projects/zooniverse/snapshot-sere...
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#19I was under the impression companies engaged in medical image diagnostics train their algos on verified clinical data (i.e for radiology the training set would have Biopsy results). Is crowd labeling valuable for medical diagnosis?
Re: Launch HN: Centaur Labs (YC W19) – Labeling Medical Images at Scale
#20Tangentially related, it kind of reminded me of Snapshot Serengeti which labels images with animals. Interestingly, these labels aren't used to train models. They're just used to filter pictures taken by automated motion sensor based cameras. The pictures with actual animals in them are used to study animal migrations. https://www.zooniverse.org/projects/zooniverse/snapshot-sere...
Yes we love zooniverse. Of course, we could do this with some non-medical tasks also. In fact, if you download the app now try out the “study break” category to classify dog breeds!