构建值得信赖的医学影像人工智能系统实现安全的临床部署.pdf

1、Building Trustworthy Medical AI Systems for Safe Clinical Deployment Crer des systmes dIA mdicale fiables pour un dploiement clinique scurisTal Arbel,PhD Canada CIFAR AI Chair,Mila Professor,McGill University,Department of Electrical and Computer Engineering Director Probabilistic Vision Group,Medic

2、al Imaging Lab Centre for Intelligent MachinesAI for Personalized Medicine:The Dream and the Challenges2Machine LearningJames DrugClinical Scenario-Current Practice 3Variety of treatments available for this patients illness Clinical Scenario-Current Practice 4Variety of treatments available for this

3、 patients illness Treatment decision:Average treatment efficacy across populationClinical Scenario Personalized Medicine5Clinical and demographic information availableTreatment decision:Average treatment efficacy conditioned on sub-group statistics 6Integrate clinical,demographic and medical images

4、into AI systemProvide clinicians with an AI tool which predicts future individual treatment response on several treatments using discovered image featuresPromise of AI for Image-Based Personalized MedicineAI SystemPromise of AI for Image-Based Personalized Medicine7Integrate clinical,demographic and

5、 medical images into AI systemProvide clinicians with an AI tool which predicts future individual treatment response on several treatments using discovered image featuresAI SystemPromise not yet met!Deep Learning Models Can Make(Potentially Deadly)Mistakes8https:/ of Interpretability of Deep Learnin

6、g Models9This patient has pleural effusionNeed to open up the black box!How did the classifier come to this conclusion?Deep Learning Models Can Be Biasedhttps:/www.pnas.org/doi/10.1073/pnas.1919012117https:/ to mitigate the biases Talk (1)*First*deep learning model for personalized medicine from pat