AI contours 21 cardiac substructures rapidly, but coronary arteries still need expert review

Automated cardiac contouring was considered acceptable with at most minor corrections for 93.8% of 630 structures, although coronary arteries remained challenging.

KEY POINTS

  • This multicenter study evaluated a RayStation-integrated deep-learning model for automatic segmentation of 21 cardiac substructures, including chambers, great vessels, coronary arteries, valves and the whole heart. The model was developed from 80 thoracic CT scans, with 70 used for training and 10 for validation. Most development scans were contrast-enhanced (78/80, 98%), and coronary artery contours incorporated planning organ-at-risk volume concepts to accommodate anatomical and motion uncertainty.
  • Independent clinical evaluation used 30 contrast-enhanced planning CT scans from two external institutions, each contributing 15 patients undergoing SBRT for central or ultracentral lung tumors. Five radiation oncologists assessed 630 automatically generated structures, including 210 coronary artery contours, using a four-point clinical acceptability scale and manual corrections when necessary.
  • Across all structures, 591 of 630 contours (93.8%) were rated as requiring no more than minor or moderate corrections by all five reviewers. 164 contours (26.0%) received unanimous scores indicating that no correction was necessary, while 39 (6.2%) were considered to require major correction by at least one clinician. No structures were entirely missing from the automated output.
  • Coronary arteries remained the most challenging anatomical group. Among 210 coronary artery contours, 186 (88.5%) were considered acceptable with at most minor or moderate corrections by all reviewers. Disagreement was more pronounced than for larger cardiac structures, with 17/210 (8.1%) receiving both no-correction and major-correction ratings from different clinicians.
  • Automated generation of all 21 structures required less than one minute per CT. Subsequent physician review and correction required a median 7.0 minutes per case (IQR 5.8–9.8), with coronary arteries generally accounting for the most demanding corrections. These times demonstrate rapid initial contour generation but do not establish total time savings against a prospectively measured fully manual workflow.
  • Median Dice similarity coefficients between the automated model and the designated reference expert were 1.00 across all structures, while clinician-corrected contours versus the reference generally achieved 0.92–1.00. However, this near-perfect model agreement is strongly dependent on the evaluation design: the reference expert participated in model development, and reviewers corrected model-generated contours rather than creating independent blinded contours. Individual coronary artery failures remained, including Dice as low as 0 for one left coronary artery contour and HD95 up to 16 mm for the middle-distal circumflex artery.
  • Most geometric comparisons showed favorable surface agreement, with mean distance to agreement generally below 1–2 mm and HD95 below 5 mm for most structures. Nevertheless, the test cohort included only 30 contrast-enhanced CT scans, without direct comparison against dedicated cardiac imaging, competing AI systems, or independent de novo contours. No analysis examined whether auto-contouring improved cardiac dosimetry, treatment decisions or clinical toxicity outcomes.

CLINICAL TAKEAWAY

Automated cardiac substructure contouring may substantially improve the practicality of detailed cardiac dose assessment in thoracic radiotherapy, especially when 21 structures can be generated within one minute. However, the reported geometric performance is influenced by the reference-contour methodology, and small coronary arteries remain vulnerable to clinically important errors; independent physician validation and dosimetric evaluation are still necessary before routine reliance on these structures.

SOURCE

Practical Radiation Oncology

Browse more research Suggest a correction