A round burgundy crest emblem with simple gold ring detailing, plain and formal in style
A wide horizontal photograph of mountain scenery near Banff with pale turquoise sky, dark evergreen slopes, and a soft muted mood
A round burgundy crest emblem mirroring the first, with plain gold ring detailing and a formal quiet mood

Annual Scientific Meeting of the Canadian Organization of Medical Physicists — July 9–12, 2014, The Banff Centre, Banff, Alberta, Canada

Automating contour segmentation in radiotherapy with machine learning

Manual contouring has long been the rate-limiting step in radiotherapy planning. A radiation oncologist or therapist can spend thirty to ninety minutes per patient drawing target volumes and organs at risk, with significant inter-observer variability. Machine learning is rapidly becoming the dominant approach to automated segmentation, compressing that effort into seconds while tightening consistency. For departments stretched on linac capacity and workforce shortages, the clinical cut-through is hard to ignore.

Australian centres have been active participants in these conversations. The annual scientific meeting hosted by the Canadian Organization of Medical Physicists remains a useful reference for anyone benchmarking workflows. Delegates preparing for credentialing can still access the CCPM examination materials shared at the conference, a useful parallel for ACPSEM TEAP candidates reviewing their own syllabus.

Evolution from atlas-based to learning-based segmentation

The first wave of commercial auto-segmentation relied on atlas-based methods, where a library of contoured cases was registered to a new patient and warped structures propagated forward. Multi-atlas approaches with label fusion improved robustness but struggled when anatomy diverged from the atlas cohort, particularly for post-surgical cavities or uncommon body habitus. Regional Australian centres, with case mix often different from the training populations of North American or European vendors, exposed these limitations early.

The shift to learning-based segmentation changed the paradigm. Rather than registering to a known case, a convolutional neural network learns where organ boundaries typically lie, drawing on hundreds or thousands of manually contoured patients. The result is a model that generalises across anatomical variation and produces a contour in under a second, even on modest hardware. The leap is comparable to moving from a paper atlas to a senior consultant with decades of cases behind them.

Deep learning architectures powering auto-contouring

Most modern auto-contouring products trace their lineage to the U-Net architecture, with its symmetric encoder-decoder and skip connections. Variants such as nnU-Net automate much of the configuration, picking preprocessing, patch size and training hyperparameters from the dataset itself. More recently, transformer-based and hybrid CNN-transformer models have entered the literature, claiming gains on small structures and on cases where boundaries are ambiguous.

Training such a model from scratch requires a curated dataset of several hundred to a few thousand contoured cases with consistent naming conventions. Few Australian centres have that volume internally, so the practical path is either a vendor-provided model or one fine-tuned on local data. Open-source checkpoints from Grand Challenge competitions and shared weights from groups at the University of Sydney and Peter MacCallum Cancer Centre have lowered the barrier, though governance and version control remain real concerns in a regulated environment.

Workflow integration in the clinic

A model that lives only on a research workstation does little for patient care. The real test is integration into the treatment planning system, whether Eclipse, RayStation, Pinnacle or a home-grown pipeline. Modern implementations push the auto-contour step automatically once simulation images are loaded, with results appearing in the structure set ready for clinician review. Sites in Brisbane and Melbourne have wired this into the booking system so contours are drafted before the radiation oncologist opens the plan.

Two operating modes dominate: hands-off, where the planner accepts the AI contour unless they see an obvious error, and review-mode, where it is treated as a first draft. Most Australian centres start in review-mode, gradually moving toward hands-off for low-risk structures such as femoral heads or contralateral lung as confidence builds. Time savings are uneven: head and neck cases, with dozens of organs at risk, benefit more than a four-field prostate.

Validation, QA and the human-in-the-loop

Regulators and professional colleges expect prospective local validation before any auto-contouring tool touches a patient. The standard metrics remain Dice similarity coefficient and Hausdorff distance, but geometric scores alone can mislead when a model consistently misses the superior aspect of a parotid by two millimetres. Visual review on stratified samples and editor time per case give a more honest picture. ARPANSA expectations and ACPSEM position papers emphasise that the clinician retains responsibility for the final contour.

A practical QA program treats each model version like new equipment. A baseline set of twenty to thirty representative cases is re-contoured after any software upgrade, scanner change or protocol revision. Drift is common, particularly as linac vendors push adaptive imaging features that subtly alter voxel statistics. Building these checks into the audit cycle, rather than treating them as a one-off commissioning task, is what separates sustainable adoption from a tool that quietly falls out of use.

Australian clinical experience and adoption

Adoption across Australia and New Zealand has been steady but uneven. Large metropolitan centres in Sydney, Melbourne and Brisbane have run prospective evaluations published in journals such as Radiotherapy and Oncology. Regional and private providers have followed, often choosing cloud-hosted products to avoid the capital cost of on-premise GPU servers. Multicentre collaborations through the Trans-Tasman Radiation Oncology Group have helped share validation datasets and standardise naming conventions between sites.

Funding models shape the rollout as much as the technology itself. Under Medicare, contouring is bundled into the planning episode rather than billed as a discrete item, so efficiency gains accrue to the provider. That calculus favours larger private networks such as Icon Group and GenesisCare, where shaving ten minutes per plan across hundreds of cases quickly justifies a licence. Public hospitals tend to require a stronger evidence base and a clearer audit trail before committing budget.

Common pitfalls and how to mitigate them

The most frequent failure mode is out-of-distribution anatomy: paediatric patients, post-mastectomy chest walls, cases with extensive dental artefact or implanted devices. The model will still produce a contour, often with high confidence, which is why a human review step cannot be quietly disabled. A second pitfall is label drift, where the auto-generated contour uses a name that does not match the departmental protocol, leading to confusion downstream when the structure is needed for plan sum or follow-up.

Data governance is a third concern Australian departments sometimes underestimate. Sending diagnostic imaging to an overseas vendor for inference may breach local privacy expectations and the Notifiable Data Breaches scheme, even when the vendor is reputable. Local processing, or a contractual arrangement with clearly defined data residency, is the safer default. Finally, complacency after a successful rollout is the silent killer: without periodic re-validation, a model trained on 2018 data may be operating on 2025 imaging that looks subtly different.

The road ahead: foundation models and adaptive therapy

The next wave is already visible. Foundation models pre-trained on millions of medical images, then fine-tuned for segmentation, are starting to outperform single-task networks, particularly on rare structures. Federated learning allows hospitals to collaboratively train a model without sharing patient data, well suited to Australia's mix of large academic centres and smaller regional sites. Coupled with online adaptive radiotherapy on platforms such as Ethos and Unity, fast auto-contouring is moving from a planning convenience to a real-time clinical necessity.

Practical recommendations for clinical implementation

  • Begin with a single anatomical site such as prostate or breast, where the task is well bounded and stakeholder buy-in is easier to secure.
  • Build a local validation set of at least thirty cases reflecting your actual patient population, including outliers that have historically challenged the team.
  • Establish baseline geometric metrics before deployment and re-measure after every software update, scanner change or major protocol revision.
  • Train physicists and radiation therapists together on the QA workflow, since both groups interact with the tool daily.
  • Engage radiation oncologists early in the evaluation, since their acceptance determines whether time savings translate into real workflow change.
  • Document a clear escalation path when the model's output is rejected, including who reviews, who annotates and how the case feeds back into future model versions.

The science keeps moving, and the gap between a research demo and a clinically trusted tool continues to narrow. For delegates and trainees planning their next conference, the icebreaker reception remains one of the best places to swap implementation stories with colleagues who have already navigated the pitfalls, and to compare notes on which products have held up under Australian conditions.