Hi everyone,
Here are some highlights of this issue. We hope you'll find them helpful.
The recent review paper from Freeman et al. examines the accuracy of AI models for detecting breast cancer in mammography screening practice. The results suggest it is unclear where AI might be most beneficial on the clinical pathway. Also, the promising results in smaller studies are not replicable in more extensive studies.
Positron emission tomography (PET) imaging is used for a wide range of abnormality detection before it shows up on other imaging tests such as CT or MRI. In recent times, deep learning has enhanced PET image quality in small patient cohorts. Through their work published at npj Digital Medicine, Chaudhari et al. have extended this ability of DL methods in the set clinical patients from multiple institutions and scanner types.
Can we predict race from medical imaging? This recent preprint from Banerjee et al. conducts a comprehensive evaluation of AI's ability to recognize patients' racial identity from medical images. They find that standard deep learning models can be trained to predict race accurately, even after adjusting potential covariates for a race, like underlying disease distribution. The authors interpret that this creates an enormous risk for the model deployment in the real world.
If you're interested in research around machine learning for drug discovery, a team led by Mila researcher Jian Tang has launched TorchDrug, an open-source interface to support rapid prototyping of drug discovery models in PyTorch.
In recent times, we have seen massive media coverage of ML research in healthcare. Recently, the BBC covered an AI research led by researchers from Cambridge University to diagnose dementia in a day. While such coverages may be helpful to advertise science to the general public, some researchers like Neil Lawrence argue that this may create unnecessary hype and may adversely affect translating machine learning into practice. What do you think?
As the world was working towards adopting AI models in the real-world clinical setting, COVID19 hit us hard, and naturally, we asked, can AI help us in this pandemic? We found the best answer in this (comic) episode of superheroes of deep learning from Zachary Lipton and Falaah Arif Khan. Don't miss it :)
Events:
ML4H: Machine Learning for Health is happening virtually on December 4, and looking for exciting papers (deadline September 13) on the intersection of healthcare and machine learning.
Research2Clinics, a NeurIPS workshop, invites articles (deadline September 17) describing innovative machine learning research focused on bridging the gap between ML research and its application in clinical practice.
NIH and the FDA join forces to lead a virtual workshop, "Next-Generation Sequencing and Radiomics: Resource Requirements for Acceleration of Clinical Applications Including AI". Broad participation is encouraged.