Leo Bman

@DanMatasic Great read and good work Dan, thanks for sharing

The skies tell us end is near... 😨 BUT antidote to tragedy is generosity 😇 Introducing RSPECT! @RSNA Pulmonary Embolism CT Dataset 𝟏𝟐𝐤 𝐂𝐓𝐬 - 𝐞𝐱𝐩𝐞𝐫𝐭 𝐥𝐚𝐛𝐞𝐥𝐬 - 𝟓 𝐬𝐢𝐭𝐞𝐬 - 𝟒 𝐜𝐨𝐮𝐧𝐭𝐫𝐢𝐞𝐬 Available now! @kaggle challenge: bourls.com/ZNhXA

@ianpanmd I'm curious to take a look & may be able to help out. thanks!

Co-editing a book -- Artificial Intelligence in Radiology: Medical and Societal Impact. We're looking for chapter authors. If you or anyone you know is interested in contributing a chapter (or several!), let me know and I'll send over the TOC!

"The medical #AI floodgates open, at a cost of $1000 per patient." My new blog post on the unexpected, important, and controversial decision for Medicare in the US to pay up to $1000 per case for a #machinelearning model to detect stroke on CT scans. lukeoakdenrayner.wordpress.com/2020/09/06/the…

FastMRI uses AI to generate MRIs from about ¼ the data as a conventional scan. See how it works: ai.facebook.com/blog/fastmri-b…

@hardmaru False. Video game skins ;)

In a global market collapse, there's no real “safe haven” Gold, bitcoin, and even cash are not safe assets.

@ben_j_radford @fchollet See @fchollet response here kaggle.com/c/abstraction-… Sounds closer to completely out-of-sample

ARC progress: ▓░░░░░░░░░░░ 6% kaggle.com/c/abstraction-…

Congratulations to these 32 teams who have advanced to Round 2 of our #STATMadness competition! Cheers to you, #Breakthru32! @kochinstitute @EmoryMedicine @Texas_Heart @UCIrvineSOM @UCSDMedSchool @BrighamWomens @StJudeResearch @MDAndersonNews @WeillCornell @UofUHealth

@NeuroStats @GaelVaroquaux @danilobzdok @KordingLab @bttyeo @f2harrell I've never run across the situation where the outcome measure was missing and I wanted to impute it, but if it was me I would only use these samples in the training data, and then check if including these observations actually improves test set performance

@GaelVaroquaux @danilobzdok @LeoBman @KordingLab @bttyeo @f2harrell Yeah, they are clearly relevant to each other, especially in deciding how to evaluate test set performance. But before the last two years or so ML literature didn't really care so much about model misspecification for the imputation model itself.

#datascience discussion I had yesterday: The goal is cross-validated prediction of continuous y from X. Before, when doing #data #imputation of missing entries in X, is supervised imputation involving y ok? If yes, why is that? @GaelVaroquaux @KordingLab @bttyeo @f2harrell

@fchollet The "cellular automata" notebook shows people are starting to think a bit outside the DL/CNN box kaggle.com/arsenynerinovs…

One interesting thing about the ARC competition is that it serves to highlight how people who use deep learning often have little idea of what deep learning actually does, and when they should be using it or not

@sarim_m @AmineKorchiMD Agreed. I think things like @EyeDiagnosis and @CaptionHealth are good examples of fitting into clinical workflow. 1st allows autonomous referral to specialist in primary care setting. 2nd could expand number of people able to take diagnostic quality images.

@AmineKorchiMD The ones that integrate into clinical workflow and save clinician time. The ones making tall claims of replacing humans won't (at least in the short to medium term)

We can easily predict that #AI companies will have a huge impact on #healthcare. It is however very hard to tell which ones or which usecases will succeed at scale.

BREAKING: FDA approves a generic version of the drug Martin Shkreli monopolized buff.ly/388Wuur

@ClaraGFaria1 @BioGoertz @AcademicChatter I've never tried matlab, but my comp psych lab uses R. R has a big community base online like python

@BioGoertz @AcademicChatter Thank you for the insight! I started with Phyton and I totally agree with you it's much more intuitive but I really wanted to get a sense of matlab since it's the language I see most coders using in #ComputationalPsychiatry which is a field I am very interested in

@AcademicChatter I wanted to ask for advice: for all the coders 💻out there what's the best/most effective way to master matlab? #GirlsWhoCode #AcademicTwitter #ComputationalPsychiatry

@GaelVaroquaux @C_Allefeld @condaforge R package installations seem to handle dependencies well 95% of the time. Moving from r to python I was surprised by how much more complicated things got, with needing separate envs via conda to keep things straight

@C_Allefeld Dependency solving comes with a very sharp computational cost (sat solvers). Dependency solving in an open world (ie not a curated set of packages such as a linux distribution) does not scale. Solutions (see @condaforge) come with tractability challenges. It's a hard problem.

Yeah, no. The versatility of #python is no excuse for not having a well designed packaging system. pip doesn't even have dependency resolver. wtf? Source: packaging.python.org/overview/

@woeaton @AcademicChatter This is an ad

Hippocampal vascularization patterns vary between individuals. Perosa et al. show that a mixed hippocampal supply, as opposed to a single one, provides a hippocampal vascular reserve that protects against cognitive impairment. bit.ly/2uhftVD

Amazing new case of a genetic drug developed for a single child w/ a unique DNA mutation. N-of-1 medicine. Told by @Erika_Check & chronicling persistence of @BradMargus, science of @timyu; commentary from @eperlste, @DrStanleyCrooke + more technologyreview.com/s/615188/dna-i…

Really interesting and well written paper on meta-learning problem: DL models sequentially trained on new classes of data "forget"/overwrite what was prev learned solution: Arch that learns to gate model updates, allowing diff parts of ntwk to specialize #AI #MachineLearning

Very cool digital phenotyping project by fellow grad student @TannerKoomar A booth that sits in @uihealthcare clinics and gathers language phenotypes