Pierre Fillard

Parlons peu, parlons biais. A🧵sur les biais en IA (avec comme d'hab des exemples tirés de la mammographie).

Not just that - it comes with a Predetermined Change Control Plan that makes it possible to claim compatibility with new vendors, as well as using our cutting-edge foudation models, without submitting a new 510k. First in the field!

And 4! 4th medical device AI that got FDA-cleared by Therapixel! 2 months after MammoScreen 3 (first AI for breast cancer detection that uses priors and does automatic reporting), we've just got the news that our Breast Density algorithm has been cleared.

We've faced intense and insightful FDA interactions, providing extensive data, making MammoScreen likely the most scrutinized AI for BC screening ever. We made it! 3rd clearance for MammoScreen! This version will elevate AI standards in breast cancer screening. Stay tuned.

Hey, MammoScreen featured in @GMA, so cool! Contributes to awareness that radiologists assisted with AI are the future of breast cancer screening. Thanks !!

@DrHughHarvey Not yet. Planned later this year. I still question the NB capabilities in having the same level of understanding of AI models than FDA. That’s my experience. Will see with MDR.

@DrHughHarvey Comparison with EU regulations has its limits. EU notified bodies are light years away from having the understanding of FDA on AI models. IMO, patient safety is much more at risk on the EU market than in the US under 510(k).

@DrHughHarvey My experience is the FDA demands rigorous training and stronger clinical evidences for AI models. It’s nearly impossible to clear a new AI model without conducting a reader study (which has limitations, but it’s another debate). Merely claiming equivalence isn’t sufficient.

Très heureux d'annoncer le démarrage d'une vaste étude prospective nationale visant à évaluer l'apport de l'IA dans le dépistage du cancer du sein ! Merci à tous les radiologues partenaires, et à @Bpifrance pour son soutien. therapixel.fr/etude-prospect…

Random thought : Le machine & plus particulièrement le deep learning doivent être la dernière option à envisager pour résoudre un problème; Uniquement quand le reste ne marche pas bien (et encore, un conseil : insistez).

10 ans déjà que tu nous as quittés. La bise, maman !

This is probably well-known in some circles but not everywhere. The most important skill for Research Scientists in AI (at least at @OpenAI) is software engineering. Background in ML research is sometimes useful, but you can usually get away with a few landmark paper.

Mon cadeau de fête des pères > votre cadeau de fête des pères

#Healthtech: 10 visionary #startups in 2022 fighting #cancer using AI. 👉techfundingnews.com/10-startups-in… #tech #news #funding #fundingnews #fintech #investment #founders #AI #Artificialintelligence #healthcare #health #top10 #innovation #technology #startup #cancer #cancercare

@alexandreabadie The man in the shadow, as always 😁

I see you @PierreFillard :)

C'est vrai, et c'est ce qui doit être notre principale préoccupation pour nous, développeurs d'IA. J'en parlais justement dans mon thread sur les biais : twitter.com/PierreFillard/… La chasse aux stratifications reste difficile mais il existe des méthodes pour s'en sortir.

incluant adresse, date de naissance, numéro de sécu, et résultat test Covid. Perso je l'ai toujours au travers de la gorge ...

On critique @doctolib (et à raison, on peut toujours mieux faire en protection des données de santé), mais on oublie rapidement que la plus grosse fuite de données de santé de ces dernières années concerne l'APHP avec 1,4M d'infos personnelles

Go train a deep learning model for that. Next stage after the chihuahua vs muffin challenge.

I am so optimistic for the future! Together with rads who make the effort to intervene in our field to understand the issues, and us, modestly listening and learning alongside them, we will make the difference. We must be together.

Having a blast @ #SBIACR2022 ! Very inspiring to talk about AI for breast cancer screening with radiologists fully understanding the potential and today’s limits of the technology without despising it. This is how we will build the future of the field, together!