A unified deep-learning framework for enhanced patient-specific quality assurance of intensity-modulated radiation therapy plans

Abstract

Background: Modern radiation therapy techniques, such as intensity-modulated radiation therapy (IMRT) and volumetric-modulated arc therapy (VMAT), use complex fluence modulation strategies to achieve optimal patient dose distribution. Ensuring their accuracy necessitates rigorous patient-specific quality assurance (PSQA), traditionally done through pretreatment measurements with detector arrays. While effective, these methods are labor-intensive and time-consuming. Independent calculation-based methods leveraging advanced dose algorithms provide a reduced workload but cannot account for machine performance during delivery.

Purpose: This study introduces a novel unified deep-learning (DL) framework to enhance PSQA. The framework can combine the strengths of measurement- and calculation-based approaches.

Methods: A comprehensive artificial training dataset, comprising 400,000 samples, was generated based on a rigorous mathematical model that describes the physical processes of radiation transport and interaction within both the medium and detector. This artificial data was used to pretrain the DL models, which were subsequently fine-tuned with a measured dataset of 400 IMRT segments to capture the machine-specific characteristics. Additional measurements of five IMRT plans were used as the unseen test dataset. Within the unified framework, a forward prediction model uses plan parameters to predict the measured dose distributions, while the backward prediction model reconstructs these parameters from actual measurements. The former enables a detailed control point (CP)-wise analysis. At the same time, the latter facilitates the reconstruction of treatment plans from the measurements and, subsequently, dose recalculation in the treatment planning system (TPS), as well as an independent second check software (VERIQA). This method has been tested with an OD 1600 SRS and an OD 1500 detector array with distinct spatial resolution and detector arrangement in combination with a dedicated upsampling model for the latter.

Results: The final models could deliver highly accurate predictions of the measurements in the forward direction and the actual delivered plan parameters in the backward direction. In the forward direction, the test plans reached median gamma passing rates better than 94% for the OD 1600 SRS measurements. The upsampled OD 1500 measurements show similar performance with similar median gamma passing rates but a slightly higher variability. The 3D gamma passing rates from the comparisons between the original and reconstructed dose distributions in patients lie between 95.4% and 98.2% for the OD 1600 SRS and 94.7% and 98.5% for the interpolated OD 1500 measurements. The dose volume histograms (DVH) of the original and the reconstructed plans, recalculated in both the TPS and VERIQA, were evaluated for the organs at risk and targets based on clinical protocols and showed no clinically relevant deviations.

Conclusions: The flexibility of the implemented model architecture allows its adaptability to other delivery techniques and measurement modalities. Its utilization also reduces the requirements of the measurement devices. The proposed unified framework could play a decisive role in automating QA workflow, especially in the context of real-time adaptive radiation therapy (ART).

Keywords: 2D arrays; artificial data; deep learning; patient‐specific quality assurance.

FIGURE 1

Schematic representation of the arrangement of the chambers forming the measurement area of (a) OD 1600 SRS array and (b) OD 1500 array. (c) Both arrays were inserted in the center of the Octavius4D phantom, which rotates synchronously with the gantry so that the array's surface is always oriented perpendicular to the beam's axis.

FIGURE 2

The proposed unified framework of PSQA that involves the implementation of a forward and a backward DL model. The forward model allows for CP‐wise comparison of the array measurements, whereas the plan parameters predicted by the backward model can be used to reconstruct the delivered RT‐plan. The reconstructed RT‐plan can then be reimported into the TPS or a second check software for dose recalculation, enabling a direct plan comparison to ease clinical decisions. The resolution of the OD 1500 measurements is increased using the upsampling model before they are evaluated within the same framework. CP, control point; DL, deep learning; PSQA, patient‐specific quality assurance; TPS, treatment planning system.

FIGURE 3

High‐level representations of the (a) forward and (b) backward model, delineated by dashed lines. The forward model takes the plan parameters, which are passed into an external encoder connected upstream of a U‐Net to predict the measurements M(x,y). The U‐Net in the backward model receives the measurement as input and passes the U‐Net output to an external decoder to predict the plan parameters for each CP(i). CP, control point.

FIGURE 4

Samples of artificial generated ${\psi }_{1,0}(x,y)$ and the corresponding measurements $M_{1600}(x,y)$ and $M_{1500}(x,y)$ of the OD 1600 SRS and OD 1500 array, respectively, derived according to Equation (4) with regular (upper row), semi‐regular (middle row), and random (lower row) MLC leaf positions.

FIGURE 5

Training scheme of the forward and backward models. The losses of the forward and backward models were calculated as loss_for and loss_back, respectively. The outputs from both models were then passed back to the corresponding models (2nd pass), after which the losses loss_for_2nd_pass and loss_back_2nd_pass were computed. The total loss was the sum of all four losses: loss_total = loss_for + loss_back + loss_for_2nd_pass + loss_back_2nd_pass.

FIGURE 6

Training and validation losses during the pretraining (300 epochs) and fine‐tuning (from 301 to 400 epochs). Left: forward and backward models; right: upsampling model.

FIGURE 7

The results from the forward and backward models of 10 selected segments from the 5 test plans (two segments per plan) measured using OD 1600 SRS.

FIGURE 8

The results from the upsampling model for four selected segments (same segments from plan 4 and plan 5 in Figure 7). The first and second columns show the raw measurement data acquired using the OD 1500 and OD 1600 SRS array, respectively. The upsampled OD 1500 measurements are shown in the third column. Profiles extracted along the designated lines are compared in the fourth columns (black circles: OD 1500, blue line: OD 1600 SRS, red line: upsampled OD 1500).

FIGURE 9

Summary of gamma passing rates as box plot for all segments in each test plan evaluated using the OD 1600 SRS (left) and upsampled OD 1500 (right) measurements.

FIGURE 10

Comparison of DVH of the original and the reconstructed plans from OD 1600 SRS as well as upsampled OD 1500 measurements for one brain (plan 3, left column) and one prostate irradiation (plan 4, right column) recalculated in Monaco TPS (upper row) and VERIQA (lower row). DVH, dose volume histograms; TPS, treatment planning system.