AI in Healthcare / AI news for Malaysia
From the archive · Source report date 16 February 2024
MOH found AI could speed up radiotherapy contouring. Clinicians still had to check it
The rapid review found large potential time savings in published studies, while drawing a hard line around clinician review, local compatibility and missing product-specific evidence.
In brief
- A Ministry of Health MaHTAS rapid review found that AI auto-contouring could sharply reduce the time needed to outline treatment targets and organs at risk on radiotherapy scans.[1]
- The review found no effectiveness, safety or cost-effectiveness evidence specifically for the proposed RT-Solution or RT-Mind products, despite finding broader evidence for AI auto-contouring.[1]
- MaHTAS said trained healthcare professionals must review and edit AI contours, and highlighted local-system compatibility, representative training data, staff training and patient explanation as implementation requirements.[1]
MaHTAS found a credible efficiency opportunity, not a finished case for autonomous use or purchase
Malaysia's health technology assessment unit examined whether artificial intelligence could assist one of radiotherapy's most painstaking planning tasks: drawing the boundaries of a tumour, treatment target and nearby organs on a patient's medical images.[1]
The answer was promising but conditional. Published studies suggested that auto-segmentation could turn parts of the contouring process from tens of minutes into seconds, yet the rapid review did not find product-specific evidence for the proposed software suite and did not remove the clinician from the loop.[1]
Contouring tells the treatment system what to target—and what to protect
Radiotherapy planning depends on accurately outlining the clinical target volume and the surrounding organs at risk on CT or other medical images. Those boundaries help a clinical team shape radiation towards a tumour while limiting unnecessary exposure to healthy structures. MaHTAS described manual delineation by clinicians as the clinical gold standard, but also as time-consuming and labour-intensive.[1]
AI auto-contouring tries to produce those outlines automatically from medical images. In the wider radiation-oncology workflow, the technology can sit between image acquisition and treatment planning, with possible uses extending into image registration, dose prediction, quality assurance and follow-up data analysis. The review focused on segmentation rather than claiming that AI could independently run the entire treatment pathway.[1]
The time-saving results were striking, but they came from different research models
The studies summarised by MaHTAS covered several cancer sites and model architectures. A prostate study reported average delineation in under 15 seconds. A lung-cancer study reported about five minutes for an automated contour compared with more than 80 minutes for complete manual delineation. A breast-cancer study reported 10.03 seconds compared with 20 to 30 minutes, while a rectal-cancer study reported 15 seconds compared with roughly 45 to 60 minutes.[1]
Those figures should not be read as one universal speed guarantee. They came from separate studies, datasets, tumour sites and technical approaches, and some studies did not report the time needed for a clinician to review and correct the result. MaHTAS also cited NICE's conclusion that AI auto-contouring with professional review was likely to be clinically equivalent to manual contouring and quicker, while more evidence was still required.[1]
The biggest gap was the proposed product itself
MaHTAS retrieved 1,107 titles in its database search, which was last run on 10 January 2023. It reported finding no evidence specifically addressing the effectiveness, safety or cost-effectiveness of RT-Solution or RT-Mind. Ten relevant studies submitted by the company were included, but these supported the broader field of AI segmentation rather than establishing the complete evidence base for the named suite.[1]
Category evidence therefore could not establish how the proposed suite would perform with Malaysian patients, systems, training and clinical workflows.[1]
Human review remained a safety requirement, not an optional extra
The review found no retrieved evidence on safety incidents, adverse events or errors caused by AI segmentation. That absence was not treated as proof of zero risk. Drawing on NICE, MaHTAS said trained healthcare professionals must always review the generated contours and edit them where necessary before use.[1]
Smaller or irregular structures could require major edits or be unusable, while atypical anatomy, previous surgery or unfamiliar imaging positions could reduce accuracy. The brief also warned that training data might underrepresent groups such as children, younger patients, the female pelvis or men with breast cancer. For those cases, manual segmentation might remain more appropriate.[1]
For Malaysia, the real test begins after the demo
MaHTAS said any AI auto-segmentation software would need to fit the local hospital ecosystem. Providers would need training packages, and hospitals would need information about the demographics behind the model's training data. The brief said models should ideally be trained on a representative national population and that patients should be informed when AI is used in their treatment planning.[1]
The economics were still uncertain. MaHTAS found no cost-effectiveness evidence for AI auto-segmentation and no exact implementation, training or software price for the proposed suite. It cited a UK range of £4 to £50 per plan and NICE's view that the technology might be cost-saving or cost-neutral, but only depending on the price and the time actually saved after review and editing.[1]
Why Malaysia should care
The practical question for Malaysian hospitals is not whether an AI contour can be generated quickly. It is whether the system works with local infrastructure, was trained on representative patients, saves time after clinician edits, and has enough safety and cost evidence for a real procurement decision.
Radiotherapy teams
AI may reduce repetitive contouring time, but the saved time must be measured after review and correction.[1]
Practical move: Pilot on representative local cases and record edit time, failure modes, dose implications and clinician acceptance.
Hospital buyers
Category-level promise is not product-level proof, and the review found no exact product cost or cost-effectiveness evidence.[1]
Practical move: Require product-specific validation, interoperability testing, training commitments and a full implementation-cost model.
Patients and families
AI can assist treatment planning, but a trained clinical professional remains responsible for reviewing the contour.[1]
Practical move: Ask the treating team how AI is used, who checks the output and how unusual anatomy is handled; this article is not medical advice.
What Malaysians can do now
- Separate a fast AI-generated contour from the final clinician-approved treatment plan when evaluating performance.
- Demand product-specific local evidence rather than relying only on studies of other models or hospitals.
- Track safety, edits, compatibility, training, patient communication and total cost during any controlled implementation.
What we still do not know
The rapid review left important product and Malaysia-specific questions unanswered.
- Whether RT-Solution or RT-Mind would achieve the published time and accuracy results in Malaysian hospitals.
- The full local price, implementation effort, training cost and cost-effectiveness after clinician review time.
- Longer-term safety outcomes and performance across representative Malaysian patient groups and difficult anatomy.
- Whether any later procurement, pilot or deployment followed the proposal reviewed by MaHTAS.
Sources
- 1.Artificial Intelligence (RT-Solution) in Auto-Segmentation for Radiotherapy Ministry of Health Malaysia, Malaysian Health Technology Assessment Section, 16 February 2024
- 2.Radixact-X9 Tomotherapy Sunway Medical Centre
- 3.Radiotherapy and Oncology Department National Cancer Institute Malaysia
