First Report On Physician Assessment and Clinical Acceptability of Custom-Retrained Artificial Intelligence Models for Clinical Target Volume and Organs-at-Risk Auto-Delineation for Postprostatectomy Patients.

PURPOSE: To assess the clinical acceptability of a commercial deep-learning-based auto-segmentation (DLAS) prostate model that was retrained using institutional data for delineation of the clinical target volume (CTV) and organs-at-risk (OARs) for postprostatectomy patients, accounting for clinical and imaging protocol variations. METHODS AND MATERIALS: CTV and OARs of 109 prostate-bed patients were used to evaluate the performance of the vendor-trained model and custom retrained DLAS models using different training quantities. Two new models for OAR structures were retrained (n = 30, 60 data sets), while separate models were trained for a new CTV structure (n = 30, 60, 90 data sets), with the remaining data sets used for testing (n = 49, 19). The dice similarity coefficient (DSC), Hausdorff distance, and mean surface distance were evaluated. Six radiation oncologists performed a qualitative evaluation scoring both preference and clinical utility for blinded structure sets. Physician consensus data sets identified from the qualitative evaluation were used toward a separate CTV model. RESULTS: Both the 30- and 60-case retrained OAR models had median DSC values between 0.91 to 0.97, improving significantly over the vendor-trained model for all OARs except the penile bulb. The brand new 60-case CTV model had a median DSC of 0.70 improving significantly over the 30-case model. DLAS (60-case model) and manual contours were blinded and evaluated by physicians with contours deemed acceptable or precise for 87% and 94% of cases for DLAS and manual delineations, respectively. DLAS-generated CTVs were scored precise or acceptable in 54% of cases, compared with the manual delineation value of 73%. The 30-case physician consensus CTV model did not show a significant difference compared with the randomly selected models. CONCLUSIONS: Custom retraining using institutional data leads to performance improvement in the clinical utility and accuracy of DLAS for postprostatectomy patients. A small number of data sets are sufficient for building an institutional site-specific DLAS OAR model, as well as for training new structures. Data indicates the workload for identifying training data sets could be shared among groups for the male pelvic region, making it accessible to clinics of all sizes.

Corrections of photon beam profiles of small fields measured with ionization chambers using a three-layer neural network.

The purpose of this work is to study the feasibility of photon beam profile deconvolution using a feedforward neural network (NN) in very small fields (down to 0.56 × 0.56 cm2 ). The method's independence of the delivery and scanning system is also investigated. Lateral beam profiles of photon fields between 0.56 × 0.56 cm2 and 4.03 × 4.03 cm2 were collected on a Siemens Artiste linear accelerator. Three scanning ionization chambers (SNC 125c, PTW 31021, and PTW 31022) of sensitive volumes ranging from 0.016 cm3 to 0.108 cm3 were used with a PTW MP3 water phantom. A reference dataset was also collected with a PTW 60019 microDiamond detector to train and test individual NNs for each ionization chamber. Further testing of the trained NNs was performed with additional test data collected on an Elekta Synergy linear accelerator using a Sun Nuclear 3D Scanner. The results were evaluated with a 1D gamma analysis (0.5 mm/0.5%). After the deconvolution, the gamma passing rates increased from 54.79% to 99.58% for the SNC 125c, from 57.09% to 99.83% for the PTW 31021, and from 91.03% to 96.36% for the PTW 31022. The delivery system, the scanning system, the scanning mode (continuous vs. step-by-step), and the electrometer had no significant influence on the results. This study successfully demonstrated the feasibility of using NN to correct the beam profiles of very small photon fields collected with ionization chambers of various sizes. Its independence of the delivery and scanning system was also shown.

Reconstruction of volume averaging effect-free continuous photon beam profiles from discrete ionization chamber array measurements using a machine learning technique.

PURPOSE: The use of the ionization chamber array ICProfiler (ICP) is limited by its relatively poor detector spatial resolution and the inherent volume averaging effect (VAE). The purpose of this work is to study the feasibility of reconstructing VAE-free continuous photon beam profiles from ICP measurements with a machine learning technique. METHODS: In- and cross-plane photon beam profiles of a 6 MV beam from an Elekta linear accelerator, ranging from 2 × 2 to 10 × 10 cm2 at 1.5 cm, 5 cm, and 10 cm depth, were measured with an ICP. The discrete measurements were interpolated with a Makima method to obtain continuous beam profiles. Artificial neural networks (ANNs) were trained to restore the penumbra of the beam profiles. Plane-specific (in- and cr-plane) ANNs and a combined ANN were separately trained. The performance of the ANNs was evaluated using the penumbra width difference (PWD, the difference between the penumbra widths of the reconstructed and the reference profile). The plane-specific and the combined ANNs were compared to study the feasibility of using a single ANN for both in- and cross-plane. RESULTS: The profiles reconstructed with all the ANNs had excellent agreement with the reference. For in-plane, the ANNs reduced the PWD from 1.6 ± 0.7 mm at 1.5 cm depth to 0.1 ± 0.1 mm, from 1.8 ± 0.6 mm at 5.0 cm depth to 0.1 ± 0.1 mm, and from 2.4 ± 0.1 mm at 10.0 cm depth to 0.0 ± 0.0 mm; for cross-plane, the ANNs reduced the PWD from 1.2 ± 0.4 mm at 1.5 cm depth, 1.2 ± 0.3 mm at 5.0 cm depth, and 1.6 ± 0.1 mm at 10.0 cm depth, to 0.1 ± 0.1 mm. CONCLUSIONS: This study demonstrated the feasibility of using simple ANNs to reconstruct VAE-free continuous photon beam profiles from discrete ICP measurements. A combined ANN can restore the penumbra of in- and cross-plane beam profiles of various fields at different depths.

The root cause analysis on failed patient-specific measurements of pencil beam scanning protons using a 2D detection array with finite size ionization chambers.

The aim of this report is to present the root cause analysis on failed patient-specific quality assurance (QA) measurements of pencil beam scanning (PBS) protons; referred to as PBS-QA measurement. A criterion to fail a PBS-QA measurement is having a <95% passing rate in a 3.0%-3.0 mm gamma index analysis. Clinically, we use a two-dimensional (2D) gamma index analysis to obtain the passing rate. The IBA MatriXX PT 2D detection array with finite size ionization chamber was utilized. A total of 2488 measurements performed in our PBS beamline were cataloged. The percentage of measurements for the sites of head/neck, breast, prostate, and other are 53.3%, 22.7%, 10.5%, and 13.5%, respectively. The measurements with a passing rate of 100 to >94%, 94 to >88%, and <88% were 93.6%, 5.6%, and 0.8%, respectively. The percentage of failed measurements with a <95% passing rate was 10.9%. After removed the user errors of either re-measurement or re-analysis, 8.1% became acceptable. We observed a feature of >3% per mm dose gradient with respect to depth on the failed measurements. We utilized a 2D/three-dimensional (3D) gamma index analysis toolkit to investigate the effect of depth dose gradient. By utilizing this 3D toolkit, 43.1% of the failed measurements were improved. A feature among measurements that remained sub-optimal after re-analysis was a sharp >3% per mm lateral dose gradient that may not be well handled using the detector size of 5.0 mm in-diameter. An analysis of the sampling of finite size detectors using one-dimensional (1D) error function showed a large dose deviation at locations of low-dose areas between two high-dose plateaus. User error, large depth dose gradient, and the effect of detector size are identified as root causes. With the mitigation of the root causes, the goals of patient-specific QA, specifically detecting actual deviation of beam delivery or identifying limitations of the dose calculation algorithm of the treatment planning system, can be directly related to failure of the PBS-QA measurements.

Evaluation of a neural network-based photon beam profile deconvolution method.

PURPOSE: The authors have previously shown the feasibility of using an artificial neural network (ANN) to eliminate the volume average effect (VAE) of scanning ionization chambers (ICs). The purpose of this work was to evaluate the method when applied to beams of different energies (6 and 10 MV) and modalities [flattened (FF) vs unflattened (FFF)], measured with ICs of various sizes. METHODS: The three-layer ANN extracted data from transverse photon beam profiles using a sliding window, and output deconvolved value corresponding to the location at the center of the window. Beam profiles of seven fields ranging from 2 × 2 to 10 × 10 cm2 at four depths (1.5, 5, 10 and 20 cm) were measured with three ICs (CC04, CC13, and FC65-P) and an EDGE diode detector for 6 MV FF and FFF. Similar data for the 10 MV FF beam was also collected with CC13 and EDGE. The EDGE-measured profiles were used as reference data to train and test the ANNs. Separate ANNs were trained by using the data of each beam energy and modality. Combined ANNs were also trained by combining data of different beam energies and/or modalities. The ANN's performance was quantified and compared by evaluating the penumbra width difference (PWD) between the deconvolved and reference profiles. RESULTS: Excellent agreement between the deconvolved and reference profiles was achieved with both separate and combined ANNs for all studied ICs, beam energies, beam modalities, and geometries. After deconvolution, the average PWD decreased from 1-3 mm to under 0.15 mm with separate ANNs and to under 0.20 mm with combined ANN. CONCLUSIONS: The ANN-based deconvolution method can be effectively applied to beams of different energies and modalities measured with ICs of various sizes. Separate ANNs yielded marginally better results than combined ANNs. An IC-specific, combined ANN can provide clinically acceptable results as long as the training data includes data of each beam energy and modality.

Error detection and classification in patient-specific IMRT QA with dual neural networks.

PURPOSE: Despite being the standard metric in patient-specific quality assurance (QA) for intensity-modulated radiotherapy (IMRT), gamma analysis has two shortcomings: (a) it lacks sensitivity to small but clinically relevant errors (b) it does not provide efficient means to classify the error sources. The purpose of this work is to propose a dual neural network method to achieve simultaneous error detection and classification in patient-specific IMRT QA. METHODS: For a pair of dose distributions, we extracted the dose difference histogram (DDH) for the low dose gradient region and two signed distance-to-agreement (sDTA) maps (one in x direction and one in y direction) for the high dose gradient region. An artificial neural network (ANN) and a convolutional neural network (CNN) were designed to analyze the DDH and the two sDTA maps, respectively. The ANN was trained to detect and classify six classes of dosimetric errors: incorrect multileaf collimator (MLC) transmission (±1%) and four types of monitor unit (MU) scaling errors (±1% and ±2%). The CNN was trained to detect and classify seven classes of spatial errors: incorrect effective source size, 1 mm MLC leaf bank overtravel or undertravel, 2 mm single MLC leaf overtravel or undertravel, and device misalignment errors (1 mm in x- or y direction). An in-house planar dose calculation software was used to simulate measurements with errors and noise introduced. Both networks were trained and validated with 13 IMRT plans (totaling 88 fields). A fivefold cross-validation technique was used to evaluate their accuracy. RESULTS: Distinct features were found in the DDH and the sDTA maps. The ANN perfectly identified all four types of MU scaling errors and the specific accuracies for the classes of no error, MLC transmission increase, MLC transmission decrease were 98.9%, 96.6%, and 94.3%, respectively. For the CNN, the largest confusion occurred between the 1-mm-MLC bank overtravel class and the 1-mm-device alignment error in x-direction class, which brought the specific accuracies down to 90.9% and 92.0%, respectively. The specific accuracy for the 2-mm-single MLC leaf undertravel class was 93.2% as it misclassified 5.7% of the class as being error free (false negative). Otherwise, the specific accuracy was above 95%. The overall accuracies across the fivefold were 98.3 ± 0.7% and 95.6% ± 1.5% for the ANN and the CNN, respectively. CONCLUSIONS: Both the DDH and the sDTA maps are suitable features for error classification in IMRT QA. The proposed dual neural network method achieved simultaneous error detection and classification with excellent accuracy. It could be used in complement with the gamma analysis to potentially shift the IMRT QA paradigm from passive pass/fail analysis to active error detection and root cause identification.