Optimization of training periods for the estimation model of three-dimensional target positions using an external respiratory surrogate

BACKGROUND: During therapeutic beam irradiation, an unvisualized three-dimensional (3D) target position should be estimated using an external surrogate with an estimation model. Training periods for the developed model with no additional imaging during beam irradiation were optimized using clinical data. METHODS: Dual-source 4D-CBCT projection data for 20 lung cancer patients were used for validation. Each patient underwent one to three scans. The actual target positions of each scan were equally divided into two equal parts: one for the modeling and the other for the validating session. A quadratic target position estimation equation was constructed during the modeling session. Various training periods for the session—i.e., modeling periods (T(M))—were employed: T(M) ∈ {5,10,15,25,35} [s]. First, the equation was used to estimate target positions in the validating session of the same scan (intra-scan estimations). Second, the equation was then used to estimate target positions in the validating session of another temporally different scan (inter-scan estimations). The baseline drift of the surrogate and target between scans was corrected. Various training periods for the baseline drift correction—i.e., correction periods (T(C)s)—were employed: T(C) ∈ {5,10,15; T(C) ≤ T(M)} [s]. Evaluations were conducted with and without the correction. The difference between the actual and estimated target positions was evaluated by the root-mean-square error (RMSE). RESULTS: The range of mean respiratory period and 3D motion amplitude of the target was 2.4–13.0 s and 2.8–34.2 mm, respectively. On intra-scan estimation, the median 3D RMSE was within 1.5–2.1 mm, supported by previous studies. On inter-scan estimation, median elapsed time between scans was 10.1 min. All T(M)s exhibited 75th percentile 3D RMSEs of 5.0–6.4 mm due to baseline drift of the surrogate and the target. After the correction, those for each T(M)s fell by 1.4–2.3 mm. The median 3D RMSE for both the 10-s T(M) and the T(C) period was 2.4 mm, which plateaued when the two training periods exceeded 10 s. CONCLUSIONS: A widely-applicable estimation model for the 3D target positions during beam irradiation was developed. The optimal T(M) and T(C) for the model were both 10 s, to allow for more than one respiratory cycle. TRIAL REGISTRATION: UMIN000014825. Registered: 11 August 2014. ELECTRONIC SUPPLEMENTARY MATERIAL: The online version of this article (10.1186/s13014-018-1019-9) contains supplementary material, which is available to authorized users.