Thank you so much for joining tonight! #RadAIchat

Another recommendation: Variability among risk classification models – perform the same at a population level, but very differently for individuals. link.springer.com/content/pd… #RadAIchat

Good study: Many studies have generalizability and bias issues, but in general AI + Human seems to be the most effective combination. mdpi.com/2075-4418/13/12/204… #RadAIchat

Important study to be aware of: RCT in Sweden on AI in screening mammography. Note different screening practices: 2D (versus mostly DBT in US), double reading and longer screening interval. thelancet.com/journals/lanon… - Dr. Milch #RadAIchat

I've been thinking AI should fully write the reports, especially for low risk/ negative exams (e.g. benign calcifications, biopsy marker, post-surgical change) and the rad just opens the study, reviews, and signs. This could also apply to US, MRI, biopsies, etc. #RadAIchat

T4 There are risk percentages associated with many of the BI-RADS descriptions, it would be cool if someday these could be automatically read from our reports and calculated. OR… keep it simple and just edit my MRI dictations! #RadAIchat

Such a great question! I think the main key here is FEWER MARKINGS. Historical CAD had too many markings which were a distraction, and ultimately led to decreased accuracy with CAD. - Dr. Milch nejm.org/doi/full/10.1056/NE… #RadAIchat

Dr Yala is the Mirai risk model expert, but this paper recently came out investigating some of the features it used to make decisions. pubs.rsna.org/doi/10.1148/ra…

As the technology proves itself, mammographers may start to convert. Making sure we educate our residents about the benefits and limitations of AI would help, and ultimately explainability would be key. #RadAIchat

2/2 Instead, we have to wait for every new static version to get approved by FDA and then implemented - Dr. Milch #RadAIchat

1/2 A big problem is that the FDA only approves AI tools that are "static" meaning they are not continually learning. A continually learning model would be more effective and more efficient. #RadAIchat

There are no prospective randomized trials demonstrating effectiveness in real-world settings, thus many practices are hesitant to adopt without high quality evidence (specifically, no trials in the US) - Milch #RadAIchat

T2 - Right now the radiologist is still responsible for the final read, so we have to trust AI and understand its strengths and weaknesses. At the same time, we need to be concerned about trusting it too much and falling into an automation bias trap.#RadAIchat

T1- Also potential for AI to assist in appropriate orders by referring physicians, streamlining patient visits, quality assessment of technologist image acquisition - Dr. Milch #RadAIchat

T1- Easy to forget that there are non-diagnostic things that would make our lives easier. Scheduling, checking on follow ups, dealing with insurance approvals. #RadAIchat

Not sure what to add to @heacockmd, but mdpi.com/2075-4418/13/12/204… Great article that summarizes some of the existing uses for AI outside mammo. #RadAIchat

Hello! I am a breast imager from UC San Diego, and have been fascinated with AI since 2017. Excited to be part of tonight's chat as your moderator. Tonight I will also be sharing Dr. Milch’s tweets from my account. #RadAIchat

Do you forsee these technologies having an impact on the training of ED or neurosurgery residents? We think a lot about overreliance on AI in rads, could that be the case for other specialists too? #RadAIchat

AI and the brain, my two favorite things. Can't wait to hear from the neurosurgery perspective tonight. Hello from San Diego! #RadAIchat

#RadAIchat can't wait for tonight's topic! I always have such a great reading list after these chats. Hello from San Diego!