The Radiation Oncology Feed For radiation oncologists · physicists · RTTs · dosimetrists
technical-relevance cervical brachytherapy deep-learning

Task-specific deep learning reduced CT metal artifacts in cervical brachytherapy

Task-specific deep learning halved pelvic artifact burden, improved applicator reconstruction to 0.1 mm, and shortened organ contouring by up to 40%.

RadOncToday · ·1 min read

KEY POINTS

  • This diagnostic development study trained four task-specific metal artifact reduction models using 500 external-beam CT scans / 2,000 slices with physics-simulated tandem-and-ovoid artifacts, then evaluated them in an independent cohort of 50 cervical brachytherapy patients. The clinical cohort included 40 standard three-channel applicators plus five cylindrical and five needle cases.
  • Task-specific optimization markedly improved all four deep-learning architectures. Mean structural similarity increased by 32.4% and peak signal-to-noise ratio by 35.7% versus the generic models (p<0.001); task-specific InDuDoNet+ performed best with SSIM 0.9736 and PSNR 36.66 dB.
  • In clinical scans, task-specific InDuDoNet+ reduced mean first-dwell-position reconstruction error to 0.10±0.18 mm, versus 0.23±0.30 mm with linear interpolation, a mean difference of −0.13 mm (95% CI −0.23 to −0.03; p<0.001). Inter-rater reconstruction reliability was excellent (ICC 0.96).
  • Artifact burden around organs at risk fell substantially: bladder artifact index decreased by 55.5% and rectal artifact index by 53.1% with task-specific InDuDoNet+ (both p<0.001), while mean CT values remained statistically unchanged, suggesting preservation of underlying tissue information.
  • Contouring geometry also improved. Bladder Hausdorff distance fell from 1.24 to 0.97 mm, while Dice similarity increased from 0.70 to 0.82 (p<0.001); comparable improvement was reported for the rectum.
  • Mean contouring time decreased from 51 to 30 seconds for bladder and from 129 to 90 seconds for rectum, corresponding to reductions of approximately 40% and 30%, respectively (p<0.001).
  • Expert image-quality scores improved from 3.5 to 4.8/5 overall with task-specific InDuDoNet+ (p<0.001). Performance generalized reasonably to cylindrical applicators, but residual artifacts remained around needle tips, highlighting weaker performance for applicator geometries not represented during training.

CLINICAL TAKEAWAY

Task-specific artifact reduction could make CT-based cervical brachytherapy faster and geometrically more reliable, particularly where stainless-steel applicators still generate substantial streak artifacts. The submillimetre reconstruction accuracy and shorter contouring times are promising, but multicentre validation and direct dose-calculation testing are needed before routine deployment.

SOURCE

Radiation Oncology