Some interesting work from Glocker et al. demonstrating the need for open weight models after investigating the risk of #Bias on closed weight models:
doi.org/10.1148/ryai...

Done!

go.bsky.app/P1zYwRs
Please share with others in the Radiation Oncology/Cancer Research space in Ireland.

This approach seems both easier and more cost-effective to implement in a clinic (compared to a full 'AI' based planning approach), as it keeps the human in the loop to oversee the process.

Education is key:

- Reviewing auto-contours is a different skill from manual delineation and requires additional training.

- Staff must maintain their manual delineation skills.

Be aware of biases:

- Automation bias can affect how clinicians/RTTs review auto-contours.

- Anchoring bias and review fatigue may lead to missed edge cases or poor system performance.

Technique and intent matter:

- Manual editing of auto-contours should only be needed if it impacts the treatment plan.

- Understanding clinical intent and treatment technique is crucial for evaluating acceptability.

When choosing a system:

- Involvement of the multidisciplinary team in evaluating and commissioning is essential.

- Commissioning should use local data representative of your clinical cohort.

- Collect common errors during commissioning for training purposes.

QA of autocontouring systems is not quite like traditional QA on a Linac:

- Interestingly, for ongoing QA day-to-day, only qualitative evaluation is recommended.

- However, broader Post-implementation monitoring is also recommended using Geometric, dosimetric, and qualitative metrics.

Auto-contouring systems should only be used as intended: staff must be aware that performance may vary across different tumour types, immobilisation equipment, and imaging modalities.