Impact of dose calculation accuracy during optimization on lung IMRT plan quality

An ultra-fast deep-learning-based dose engine for prostate VMAT via knowledge distillation framework with limited patient data

Objective. Deep-learning (DL)-based dose engines have been developed to alleviate the intrinsic compromise between the calculation accuracy and efficiency of the traditional dose calculation algorithms. However, current DL-based engines typically possess high computational complexity and require powerful computing devices. Therefore, to mitigate their computational burdens and broaden their applicability to a clinical setting where resource-limited devices are available, we proposed a compact dose engine via knowledge distillation (KD) framework that offers an ultra-fast calculation speed with high accuracy for prostate Volumetric Modulated Arc Therapy (VMAT).Approach. The KD framework contains two sub-models: a large pre-trained teacher and a small to-be-trained student. The student receives knowledge transferred from the teacher for better generalization. The trained student serves as the final engine for dose calculation. The model input is patient computed tomography and VMAT dose in water, and the output is DL-calculated patient dose. The ground-truth \dose was computed by the Monte Carlo module of the Monaco treatment planning system. Twenty and ten prostate cases were included for model training and assessment, respectively. The model's performance (teacher/student/student-only) was evaluated by Gamma analysis and inference efficiency.Main results. The dosimetric comparisons (input/DL-calculated/ground-truth doses) suggest that the proposed engine can effectively convert low-accuracy doses in water to high-accuracy patient doses. The Gamma passing rate (2%/2 mm, 10% threshold) between the DL-calculated and ground-truth doses was 98.64 ± 0.62% (teacher), 98.13 ± 0.76% (student), and 96.95 ± 1.02% (student-only). The inference time was 16 milliseconds (teacher) and 11 milliseconds (student/student-only) using a graphics processing unit device, while it was 936 milliseconds (teacher) and 374 milliseconds (student/student-only) using a central processing unit device.Significance. With the KD framework, a compact dose engine can achieve comparable accuracy to that of a larger one. Its compact size reduces the computational burdens and computing device requirements, and thus such an engine can be more clinically applicable.

Impact of the dose quantity used in MV photon optimization on dose distribution, robustness, and complexity

PURPOSE

Convolution/superposition algorithms used in megavoltage (MV) photon radiotherapy model radiation transport in water, yielding dose to water-in-water (D_w,w ). Advanced algorithms constitute a step forward, but their dose distributions in terms of dose to medium-in-medium (D_m,m ) or dose to water-in-medium (D_w,m ) can be problematic when used in plan optimization due to their different dose responses to some atomic composition heterogeneities. Failure to take this into account can lead to undesired overcorrections and thus to unnoticed suboptimal and unrobust plans. Dose to reference-like medium (D_ref,m* ) was recently introduced to overcome these limitations while ensuring accurate transport. This work evaluates and compares the performance of these four dose quantities in planning target volume (PTV)-based optimization.

METHODS

We considered three cases with heterogeneities inside the PTV: virtual phantom with water surrounded by bone; head and neck; and lung. These cases were planned with volumetric modulated arc therapy (VMAT) technique, optimizing with the same setup and objectives for each dose quantity. We used different algorithms of the Varian Eclipse treatment planning system (TPS): Acuros XB (AXB) for D_m,m and D_w,m , and Analytical Anisotropic Algorithm (AAA) for D_w,w . D_ref,m* was obtained from D_m,m distributions using an in-house software considering water as the reference medium (D_w,m* ). The optimization process consisted of: (1) common first optimization, (2) dose distribution computed for each quantity, (3) re-optimization, and (4) final calculation for each dose quantity. The dose distribution, robustness to patient setup errors, and complexity of the plans were analyzed and compared.

RESULTS

The quantities showed similar dose distributions after the optimization but differed in terms of plan robustness. The cases with soft tissue and high-density heterogeneities followed the same pattern. For AXB D_m,m , cold regions appeared in the heterogeneities after the first optimization. They were compensated in the second optimization through local fluence increases, but any positional mismatch impacted robustness, with clinical target volume (CTV) variations from the nominal scenario around +3% for bone and up to +7% for metal. For AXB D_w,m the pattern was inverse (hot regions compensated by fluence decreases) and more pronounced, with CTV dose variations around -7% for bone and up to -17% for metal. Neither AXB D_w,m* nor AAA D_w,w presented these dose inhomogeneities, which resulted in more robust plans. However, D_w,w differed markedly from the other quantities in the lung case because of its lower radiation transport accuracy. AXB D_m,m was the most complex of the four dose quantities and AXB D_w,m* the least complex, though we observed no major differences in this regard.

CONCLUSIONS

The dose quantity used in MV photon optimization can affect plan robustness. D_w,w distributions from convolution/superposition algorithms are robust but may not provide sufficient radiation transport accuracy in some cases. D_m,m and D_w,m from advanced algorithms can compromise robustness because their different responses to some composition heterogeneities introduce additional fluence compensations. D_ref,m* offers advantages in plan optimization and evaluation, producing accurate and robust plans without increasing complexity. D_ref,m* can be easily implemented as a built-in feature of the TPS and can facilitate and simplify the treatment planning process when using advanced algorithms. Final reporting can be kept in D_m,m or D_w,m for clinical correlations.

Boosting radiotherapy dose calculation accuracy with deep learning

In radiotherapy, a trade‐off exists between computational workload/speed and dose calculation accuracy. Calculation methods like pencil‐beam convolution can be much faster than Monte‐Carlo methods, but less accurate. The dose difference, mostly caused by inhomogeneities and electronic disequilibrium, is highly correlated with the dose distribution and the underlying anatomical tissue density. We hypothesize that a conversion scheme can be established to boost low‐accuracy doses to high‐accuracy, using intensity information obtained from computed tomography (CT) images. A deep learning‐driven framework was developed to test the hypothesis by converting between two commercially available dose calculation methods: Anisotropic analytic algorithm (AAA) and Acuros XB (AXB). A hierarchically dense U‐Net model was developed to boost the accuracy of AAA dose toward the AXB level. The network contained multiple layers of varying feature sizes to learn their dose differences, in relationship to CT, both locally and globally. Anisotropic analytic algorithm and AXB doses were calculated in pairs for 120 lung radiotherapy plans covering various treatment techniques, beam energies, tumor locations, and dose levels. For each case, the CT and the AAA dose were used as the input and the AXB dose as the “ground‐truth” output, to train and test the model. The mean squared errors (MSEs) and gamma passing rates (2 mm/2% & 1 mm/1%) were calculated between the boosted AAA doses and the “ground‐truth” AXB doses. The boosted AAA doses demonstrated substantially improved match to the “ground‐truth” AXB doses, with average (± s.d.) gamma passing rate (1 mm/1%) 97.6% (±2.4%) compared to 87.8% (±9.0%) of the original AAA doses. The corresponding average MSE was 0.11(±0.05) vs 0.31(±0.21). Deep learning is able to capture the differences between dose calculation algorithms to boost the low‐accuracy algorithms. By combining a less accurate dose calculation algorithm with a trained deep learning model, dose calculation can potentially achieve both high accuracy and efficiency.

On the pitfalls of PTV in lung SBRT using type-B dose engine: an analysis of PTV and worst case scenario concepts for treatment plan optimization

Background

PTV concept is presumed to introduce excessive and inconsistent GTV dose in lung stereotactic body radiotherapy (SBRT). That GTV median dose prescription (D₅₀) and robust optimization are viable PTV–free solution (ICRU 91 report) to harmonize the GTV dose was investigated by comparisons with PTV–based SBRT plans.

Methods

Thirteen SBRT plans were optimized for 54 Gy / 3 fractions and prescribed (i) to 95% of the PTV (D₉₅) expanded 5 mm from the ITV on the averaged intensity project (AIP) CT, i.e., PTV_ITV, (ii) to D₉₅ of PTV derived from the van Herk (VH)‘s margin recipe on the mid–ventilation (MidV)–CT, i.e., PTV_VH, (iii) to ITV D₉₈ by worst case scenario (WCS) optimization on AIP,i.e., WCS_ITV and (iv) to GTV D₉₈ by WCS using all 4DCT images, i.e., WCS_GTV. These plans were subsequently recalculated on all 4DCT images and deformably summed on the MidV–CT. The dose differences between these plans were compared for the GTV and selected normal organs by the Friedman tests while the variability was compared by the Levene’s tests. The phase–to–phase changes of GTV dose through the respiration were assessed as an indirect measure of the possible increase of photon fluence owing to the type–B dose engine. Finally, all plans were renormalized to GTV D₅₀ and all the dosimetric analyses were repeated to assess the relative influences of the SBRT planning concept and prescription method on the variability of target dose.

Results

By coverage prescriptions (i) to (iv), significantly smaller chest wall volume receiving ≥30 Gy (CW_V30) and normal lung ≥20 Gy (NL_V20Gy) were achieved by WCS_ITV and WCS_GTV compared to PTV_ITV and PTV_VH (p > 0.05). These plans differed significantly in the recalculated and summed GTV D₂, D₅₀ and D₉₈ (p <  0.05). The inter–patient variability of all GTV dose parameters is however equal between these plans (Levene’s tests; p > 0.05). Renormalizing these plans to GTV D₅₀ reduces their differences in GTV D₂, and D₉₈ to insignificant level (p > 0.05) and their inter–patient variability of all GTV dose parameters. None of these plans showed significant differences in GTV D₂, D₅₀ and D₉₈ between respiratory phases, nor their inter–phase variability is significant.

Conclusion

Inconsistent GTV dose is not unique to PTV concept but occurs to other PTV–free concept in lung SBRT. GTV D₅₀ renormalization effectively harmonizes the target dose among patients and SBRT concepts of geometric uncertainty management.

Advanced dose calculation algorithms in lung cancer radiotherapy: Implications for SBRT and locally advanced disease in deep inspiration breath hold

PURPOSE

Evaluating performance of modern dose calculation algorithms in SBRT and locally advanced lung cancer radiotherapy in free breathing (FB) and deep inspiration breath hold (DIBH).

METHODS

For 17 patients with early stage and 17 with locally advanced lung cancer, a plan in FB and in DIBH were generated with Anisotropic Analytical Algorithm (AAA). Plans for early stage were 3D-conformal SBRT, 45 Gy in 3 fractions, prescribed to 95% isodose covering 95% of PTV and aiming for 140% dose centrally in the tumour. Locally advanced plans were volumetric modulated arc therapy, 66 Gy in 33 fractions, prescribed to mean PTV dose. Calculation grid size was 1 mm for SBRT and 2.5 mm for locally advanced plans. All plans were recalculated with AcurosXB with same MU as in AAA, for comparison on target coverage and dose to risk organs.

RESULTS

Lung volume increased in DIBH, resulting in decreased lung density (6% for early and 13% for locally-advanced group). In SBRT, AAA overestimated mean and near-minimum PTV dose (p-values < 0.01) compared to AcurosXB, with largest impact in DIBH (differences of up to 11 Gy). These clinically relevant differences may be a combination of small targets and large dose gradients within the PTV. In locally advanced group, AAA overestimated mean GTV, CTV and PTV doses by median less than 0.8 Gy and near-minimum doses by median 0.4-2.7 Gy. No clinically meaningful difference was observed for lung and heart dose metrics between the algorithms, for both FB and DIBH.

CONCLUSIONS

AAA overestimated target coverage compared to AcurosXB, especially in DIBH for SBRT.

Investigating the dosimetric effects of grid size on dose calculation accuracy using volumetric modulated arc therapy in spine stereotactic radiosurgery

PURPOSE

Sharp dose gradients between the target and the spinal cord are critical to achieve dose constraints in spine stereotactic radiosurgery (SRS), however the accuracy of the doses to the spinal cord at these high dose gradients is sensitive to the how the dose is sampled across the structure using a discretized isotropic calculation grid. In this study, the effect of the grid size (GS) on the dosimetric accuracy of volumetric modulated arc therapy (VMAT) spine SRS plans was investigated.

METHODS

The Eclipse v11.0 Anisotropic Analytical Algorithm (AAA) algorithm was used for dose calculation. Plan qualities of fifty treatment plans were evaluated with a GS of 2.5 (AAA's default value), 1.5 and 1mm. All plans were prescribed to the 90% isodose line in 1 fraction. Parameters used for plan comparison included the distance-to-fall-off (DTF) between the 90% and 50% isodose levels in the axial plane, planning tumor volume (PTV) coverage to 99%, 95%, 5% and 0.03cc, dose to 10% (Cord_D10%) and 0.03cc (Cord_D0.03cc) of the spinal cord sub volume. The dosimetric accuracy was evaluated based on film dosimetry percent gamma pass rate, line profile through the cord. Calculation times between different grid sizes as well as DVH algorithm differences between two treatment planning systems (Eclipse vs Velocity) were compared. Paired t-test was used to investigate the statistical significance.

RESULTS

The DTF decreased for all plans with 1mm compared to 1.5mm and 2.5mm GS (2.52±0.54mm, 2.83±0.58mm, 3.30±0.64, p<0.001). Relative to the 1mm GS, Cord_D0.03cc and Cord_D10% increased by 6.24% and 7.81% with the 1.5mm GS, and 9.80% and 13% with the 2.5mm GS. Film analysis demonstrated higher gamma pass rates for 1.5mm GS compared to 1 and 2.5mm GS (95.9%±5.4%, 94.3%±6.0%, 93.6%±5.4%, p<0.001), however 1mm GS showed better agreement in the high dose gradient near the cord. Calculation times for 1mm GS plans increased for 1.5 and 2.5mm GS (61% and 84%, p<0.001). The average difference between the two treatment planning systems was approximately 0-1.2%. A maximum difference of 5.9% occurred for Cord_D0.03cc for the 1mm GS.

CONCLUSION

Plans calculated with a 1mm grid size resulted in the most accurate representation of the dose delivered to the cord, however resulted in less uniform dose distributions in the high dose region of the PTV. The use of a 1.5mm grid size may balance accurate cord dose and PTV coverage, while also being more practical with respect to computation time.

Development and evaluation of a clinical model for lung cancer patients using stereotactic body radiotherapy (SBRT) within a knowledge-based algorithm for treatment planning

The purpose of this study was to describe the development of a clinical model for lung cancer patients treated with stereotactic body radiotherapy (SBRT) within a knowledge‐based algorithm for treatment planning, and to evaluate the model performance and applicability to different planning techniques, tumor locations, and beam arrangements. 105 SBRT plans for lung cancer patients previously treated at our institution were included in the development of the knowledge‐based model (KBM). The KBM was trained with a combination of IMRT, VMAT, and 3D CRT techniques. Model performance was validated with 25 cases, for both IMRT and VMAT. The full KBM encompassed lesions located centrally vs. peripherally (43:62), upper vs. lower (62:43), and anterior vs. posterior (60:45). Four separate sub‐KBMs were created based on tumor location. Results were compared with the full KBM to evaluate its robustness. Beam templates were used in conjunction with the optimizer to evaluate the model's ability to handle suboptimal beam placements. Dose differences to organs‐at‐risk (OAR) were evaluated between the plans generated by each KBM. Knowledge‐based plans (KBPs) were comparable to clinical plans with respect to target conformity and OAR doses. The KBPs resulted in a lower maximum spinal cord dose by 1.0±1.6Gy compared to clinical plans, p=0.007. Sub‐KBMs split according to tumor location did not produce significantly better DVH estimates compared to the full KBM. For central lesions, compared to the full KBM, the peripheral sub‐KBM resulted in lower dose to 0.035 cc and 5 cc of the esophagus, both by 0.4Gy±0.8Gy, p=0.025. For all lesions, compared to the full KBM, the posterior sub‐KBM resulted in higher dose to 0.035 cc, 0.35 cc, and 1.2 cc of the spinal cord by 0.2±0.4Gy, p=0.01. Plans using template beam arrangements met target and OAR criteria, with an increase noted in maximum heart dose (1.2±2.2Gy, p=0.01) and GI (0.2±0.4, p=0.01) for the nine‐field plans relative to KBPs planned with custom beam angles. A knowledge‐based model for lung SBRT consisting of multiple treatment modalities and lesion locations produced comparable plan quality to clinical plans. With proper training and validation, a robust KBM can be created that encompasses both IMRT and VMAT techniques, as well as different lesion locations.

PACS number(s): 87.55de, 87.55kh, 87.53Ly