Wrong focus (Score 1)

Automated general image interpretation is a fantasy currently. Diagnosis systems are where we should be focusing. AI folk consider radiological interpretation as some massive image recognition task, a giant CIFAR challenge, when in reality it is nothing like this. Even the simplest of medical imaging, the humble chest x-ray (CXR), requires near human level AI for useful interpretation. We can train CNN's etc. to recognise possible lesions yet its much harder to know which ones are important. Sensitivity is easy. Specificity is the challenge. Specificity is what the human brings. Outside of focused applications such as mammography we have nowhere near the required datasets to train models for useful specificity with current algorithms. The range of questions a useful system needs to answer is too broad (is pathology present?, has there been interval improvement?, characterise that lesion, where is that drain sitting?, ...). Even for the simple CXR we would end up requiring many separate models. Now we have a new problem of deciding which ones to apply and when. One day we will figure this out but for now there are far lower hanging fruit to tackle such as standardising and optimising the diagnostic pathways that request imaging. For example when you are sick it is unlikely that you are the first person in history with a particular disease state thus diagnosis is initially a search problem within a finite space. Only after this search is complete should patient pathways start to diverge significantly. Why then does the same patient presentation trigger a different initial response depending who they see? Medicine should be consistent, the idea of a "second opinion" is terrifying. The space is too big for a single individual to cover so each has their own little corner that becomes their "hammer" for every "nail" they see. Too much reliance on "hard won gestalt". There is truth to the expression: "your gut is full of shit". A virtual radiologist is a quantum leap away yet, but we could beat the diagnostic pants off the average physician right now with existing AI approaches. Making this happen is where AI should be spending its time. Creating an environment where it can happen is where government should be spending its time. This is how we still have a health system in 20 years when their aren't enough taxpayers to cover the current level of diagnostic inefficiency.