A DICOM standard has been developed to report common data elements such as PSA, ethnicity, age, etc in prostate MRI image metadata. This is one approach could provide a universal standard for unbiased data collection #AJRChat

Q5: Developing expertise in med imaging in this field including diversity, inclusion... Collecting standard data on all cases that are potential confounders such as known risk profile information and reporting it... #AJRChat

Q4: The ACR DSI @AcrDsi is taking some initiatives to provide platforms for validation which should help move the field forward #AJRChat

Q4 Agree 100% with Baris. In my view use of public data sets, the need for open source code and evidence of stability after deployment are not emphasized enough in CLAIM for those in clinical practice. #AJRChat

Q3 Quality assurance is definitely up there. There is promising work coming from the manufacturers and places like NIH with Baris @radiolobt and elsewhere for PIRADS classification which is one of the "holy grails" #AJRChat

Q3: We can speculate on things to hit us first in prostate, my vote is for noise reduction with pulse sequences (DWI) and prostate volumetrics. Others still need work but are likely to come. #AJRChat

Q3: Categories: At the scanner-noise reduction, quality assurance; At the workstation: gathering history from EPR, cancer localization/segmentation, report generation #AJRChat

Q2: Menial task examples: prostate volume, change in prostate vol over time. Many others outside prostate domain. Co-registration for detecting changes over time #AJRChat

Q2: Failure: A prostate cancer DL semantic segmentation tool that completely missing a cancer that fills the whole gland on MRI #AJRChat

Q2 - Advantage DL does not fatigue. Measuring prostate volumes accurately is a concrete example of a DL assisted functionality we could use for PSA density #AJRChat

Q2: You wouldn't leave a 2 year old alone in the house. Human supervision required #AJRChat

Q2: Independent DL without human supervision in many domains is still difficult. When DL fails it often does so catastrophically. #AJRChat

Q2: DL has as yet unrealized potential. Advantage - it has the potential to free us from menial task, improve our efficency, augment us #AJRChat

Q1: Machine Learning is an algorithmic approach where the algorithm learns and adapts from drawing inferences from data. It is data driven and encompasses many approaches #AJRChat

Q1: Typically deep means many layers with most successful networks have 16 or more with 50 not being unusual #AJRChat

Deep Learning is a type of ML. It is based on artificial neural networks. It has an input layer and an output layer. The deep part refers to having many intermediate layers. #AJRChat

Hi. Glad to be with you today #AJRChat