Short notes on radiology and AI
There’s heated discussion on the impact of AI tools (or machine learning to be more accurate) on radiology. Here’s a points I summarised from various conferences and readings on the internet.
The problem AI solves
- A lot of reports, but not enough radiologists. AI services on the cloud can read scans much faster, more consistent, more objective and likely cheaper.
- However, some argues that the most valuable information radiologists provide is not the diagnosis, but to help answer the questions from the clinical team.
Issues with medical data
- Volume: medical imaging data are abundant compared to other specialties, but it’s still small compared to general data such as road data for automated vehicles, everyday images such as ImageNet.
- Privacy: already a huge hurdle for many research projects, but another issue lies in the fact that current anonymisation techniques may not be sufficient.
- Source and bias: although we have many promising results from ML with the current data sets, there are always issues with “rubbish in, rubbish out”.
- Labels and ground truth. The machine learning models cannot diagnose better than human when because the datasets were labelled by humans. This statement does have caveats, such as the machine may perform better than trainees or be more consistent.
- Rare cases: we simply don’t have enough of rare medical cases to train the machine
A few issues with AI-generated results
- Clinical significance: this is where clinical experience is important.
- Explainability: currently a rather hot field.
- Generalisation: the training data may be biased and the result may not be generalisable to other populations.
- False positives: even radiologists may disagree on the final diagnosis.
Future of radiology
- Not only making diagnosis, but guiding treatment
- Participate in relevant research
Future of medical innovation
:PROPERTIES: nil:END:
- Workflow improvement rather than systems innovation. The real pain point in most hospitals is actually in its workflow process. Tools that make this more efficient is likely to make more reallife improvement to patients.
- Basic infrastructure improvements will offer further higher yield: fax machines, electronic health records