KEY POINTS
- The DART workflow was retrospectively evaluated in 12 head and neck patients treated at Austin Health from 2020–2023 who had already undergone clinically indicated mid-treatment replanning. Seven were treated with IMRT and five with VMAT.
- Clinical plans used a 5-mm CTV-to-PTV margin with three typical dose levels—70 Gy, 63 Gy and 56 Gy in 35 fractions. Replanning occurred between weeks 1 and 5, most commonly during weeks 3–4.
- DART starts from treatment CBCT and creates a synthetic CT using image enhancement, separate head/body rigid registrations, image stitching and deformable registration. This “double-fusion” strategy was specifically designed to distinguish changes such as neck tilt from weight loss, rather than trying to capture both with a single registration.
- Geometric agreement with the clinical replan CT was good: average Dice coefficients were >80% for all evaluated structures except the submandibular glands (>75%), excluding the very small cochlear structures. Mean distance-to-agreement averaged <2 mm, and all values were <3 mm.
- Dosimetric agreement was similarly strong. Average DVH differences between synthetic and replan CT calculations were within 1%, all evaluated dose differences were within 3%, and total gamma passing rates were >95%.
- Using the clinical decision to replan as ground truth and a unified 2-Gy change threshold across target and organ-at-risk DVH metrics, DART achieved 92% sensitivity and 92% specificity for identifying clinically relevant adaptation.
- Organ-at-risk changes—particularly involving the salivary glands—were frequent drivers of adaptation flags. Four of 12 patients also showed relevant deterioration in D98 of the elective low-dose target, whereas only two had major changes affecting the high-dose target, highlighting why adaptation assessment cannot rely on target coverage alone.
CLINICAL TAKEAWAY
DART offers a practical route from routine CBCT to quantitative “does this patient actually need replanning?” decision support without automatically rescanning everyone. The 92%/92% performance is encouraging, but with only 12 preselected patients who all underwent clinical replanning, larger prospective validation is essential before using the threshold as an autonomous adaptation trigger.