A review of radiotherapy linear accelerator quality control techniques

Conventional

Historically, the basis of a QC program is built on testing individual linac components in isolation using test methods, frequencies, and tolerances¹⁴ that are not specific to the linac manufacturer, age, or workload. These are typically prescribed by regulatory and professional bodies via guidance documents and reports, such as the those published by the Institute of Physics and Medicine (IPEM),²⁶ the American Association of Physicists in Medicine (AAPM),^27–30 the Netherlands Commission for Radiation Dosimetry (NCS),^{31,  32} and Australasian College of Physical Scientists and Engineers in Medicine (ACPSEM).³³ An aim of these guidance documents is to standardize practice; however, a 2024 UK survey of linac QC³⁴ demonstrated a variability in implementation of QC practices, despite 91% of responding centers stating that they used IPEM Report 81 (2nd edition, 2018)²⁶ to inform their QC schedule.³⁴ In addition, the rate of publication of official guidance can struggle to keep pace with rapidly advancing radiotherapy techniques.⁵ Furthermore, Barnes¹³ describes how some IPEM²⁶ and AAPM^{27,  29} reports mainly detail static testing of linac components and questions their validity of testing modern dynamic radiotherapy. Barnes¹³ also states that while the NCS reports^{31,  32} detail more dynamic testing protocols, it is unclear if these methods have become widespread.¹³

Conventional tolerances are defined thresholds that, when used simplistically, require a performance parameter to deteriorate past a limit before a fault is identified. This can result in binary decision-making, and often does not consider the variability within the threshold which could indicate deteriorating performance prior to breakdown.^{5,  35}

Conventional QC frequencies are fixed time intervals and do not consider the age or workload of treatment machines, which could impact upon the required safe testing frequency.³⁶ A conventional technique is often tedious and time-consuming due to the component-based nature and manual methods employed,^{15,  16} and does not consider the failure modes that exist when components are combined as in treatment delivery.^{5,  14}

Automated and manufacture integrated

Manufacturer integrated QC (MIQC) is defined by Pearson et al¹⁷ as “An automated quality control (QC) tool supplied by the treatment machine vendor.”¹⁷ MIQC utilizes existing linac components and data, such as the electronic portal imaging device (EPID) and machine log files, to automatically perform and analyze QC tests. The results of a 2022 IPEM survey into the use of MIQC¹⁷ stated that 25% of survey respondents envisaged the potential future scope of MIQC as “detecting errors ahead of time,” 21% as “alternative to daily QA,” and 21% “reduce the burden on monthly QC.”¹⁷ The follow-up IPEM topical report: guidance on the use of MIQC¹⁸ stated that benefits may include “reduced error, standardization, and big-data analysis.”¹⁸

Despite these benefits, there is still uncertainty as to whether MIQC can be solely relied upon to verify machine performance. Multiple studies have compared MIQC methods to conventional methods.^37–41 The use of an EPID may not be a sufficient replacement for output measurements due to uncertainties in dose measurement compared to conventional methods.^{38,  41–43} To reduce this uncertainty, several guidance documents recommend comparison of chamber and Varian Machine Performance Check (MPC) (Varian Medical Systems, Palo Alto, CA, USA) output as a routine QC test.^{26,  34,  38,  41}

The independence of MIQC to the delivery device being tested has also been raised as a possible concern.^{17,  18} A total of 81% of the 2022 IPEM survey respondents considered EPID measurements to meet the threshold for independence as the EPID panel is independent of the treatment beam parameters it tests.¹⁷ This principle, given in the guidance report by Pearson et al,¹⁸ defines 3 aspects of independence: financial, manufacturer, and technical. For MIQC to be implemented technical independence should be proven by the user. In addition, an overlap period of performing both MIQC and conventional QC is recommended.¹⁸

Third-party automated QC has been assessed in a number of studies.^43–45 Bonanno et al⁴⁵ compared DoseLab AutoQA module (Varian Medical Systems, Palo Alto, CA, USA) to conventional QC methods for monthly routine QC. Potential time savings were among the parameters evaluated. DoseLab took 80 minutes to complete monthly QC compared to 195 minutes for the conventional technique. The time savings were mainly seen in the analysis, only 20 minutes of machine time was saved compared to the 115 minutes overall time saving. Stambaugh et al⁴³ compared the accuracy and efficiency of SunCHECK Machine (Sun Nuclear Corporation, Melbourne, FL, USA) with Varian MPC for daily QC. Output measurements were also compared to measurements acquired by their monthly standard, the IC profiler (Sun Nuclear Corporation, Melbourne, FL, USA). Stambaugh et al⁴³ found that SunCHECK machine tests took on average 22 minutes compared to 15 minutes with MPC. MPC output saw a drift in agreement with SunCHECK of 1.4% over 3 months, whereas SunCHECK drifted in line with the IC profiler. MPC drift has been found in other studies^{38,  41,  42} and as a result, regular re-baselining is recommended to maintain confidence in the results.⁴² This approach has been taken in the UK where 21% of centers perform monthly comparisons between MPC and ion chamber to monitor for drift.³⁴ SunCHECK machine was stated as being preferred as it could cover more AAPM TG-142 recommendations²⁷ than MPC, yet MPC was deemed a useful addition as it provides valuable technical information for service engineers.⁴³

This use of machine log files as a method of machine-reported performance is also regularly investigated. Log files can be used to compare intended and actual parameters during delivery.^46–48 Son et al⁴⁹ used comparison of multi-leaf collimator (MLC) and monitor unit (MU) log files to determine if their Halcyon performance was in compliance with AAPM TG 142²⁷ guidelines. As a result of this analysis, Son et al⁴⁹ determined log file analysis could be used to test linac performance parameters and suggested further work comparing log files with other QC methods (eg, EPID or phantom measurements) should be carried out. Wojtasik et al⁵⁰ performed multivariate analysis on MLC log files to predict MLC failure before it occurred. Their model was found to have a 62% MLC failure prediction sensitivity, which allowed for prospective maintenance. The model was not able to identify which leaf required replacement. Many other studies employ the use of both log files and deep learning methods to predict MLC positions^51–53 (see Artificial Intelligence section for further detail).

Risk

Risk-based methods evaluate a whole process and determine which QC should be performed to reduce the highest rated risks in a system. AAPM TG-100⁵⁴ gives advice on the application of risk analysis to radiotherapy, specifically failure modes effects analysis (FMEA). Liu et al⁵⁵ conducted a literature review of the use of FMEA for proactive healthcare risk assessment. While the review was not specific to radiotherapy, the majority of results returned were in this field. In addition, Liu et al⁵⁵ states that FMEA in healthcare should consider more risk factors in addition to the conventional 3 (occurrence [O], severity [S], and detectability [D]). This was done by Ochi et al⁵⁶ when evaluating the frequency and tolerances of standard outputs. An additional 2 indices were added: severity to output drifting and number of treatment fractionations. Different frequency and tolerance combinations were used to calculate Risk Priority Numbers (RPNs) and quantitatively determine the optimal combination for either safety or efficiency. Conversely, Rah et al⁵⁷ recommends the use of a modified healthcare FMEA (mFMEA) using 2 indices (severity and frequency) as it was found to be conceptually easier to implement. Kawahara et al⁵⁸ used 3 indices, but these differed from the conventional O, S, and D. Instead, the indices used were biological effect, occurrence, and cost of labor and repairs to assess the impact of treatment interruption due to machine breakdown in terms of biological effect to the patient and the economic cost to the institution. They found that MLC faults had the highest biological risk and laser guard the highest economic risk.

O’Daniel and Yin⁵⁹ used FMEA to determine clinic-specific linear accelerator QC test frequencies for the tests detailed in AAPM TG-142.²⁷ Occurrence was determined from 9 years of linac QC results, severity was determined through introduced errors in head and neck plans, and detectability was set to be equal for all TG-142 tests considered. This method claimed to reduce bias in comparison to a method based on expert opinion. The greatest severity was found for daily imaging QC and the greatest occurrence for output and laser QC. Consequently, output and laser QC was conducted daily and imaging QC weekly, reducing daily QC time by 5-10 minutes. Bonfantini et al⁶⁰ used both quantitative methods, via QC records, and expert opinions, from surveys, to determine risk scores to prioritize linac QC. This allowed for reduced bias while maintaining staff experience. Pearson et al⁶¹ used a combination of QC records and scoring surveys to validate a hybrid QC approach of conventional QC and MIQC. In this case MPC, daily QA3 (Sun Nuclear Corporation, Melbourne, FL, USA), and Delta4 (ScandiDos AB, Uppsala, Sweden) were used. FMEA helped determine that a hybrid approach would be appropriate, but more MLC QC tests would be beneficial.

De Kerf et al⁶² used FMEA to assess the risk of a fully automated workflow (FAW). The highest risk found was staff interaction with the FAW, for example, poor judgment of the outcome and a deskilled workforce. This was of particular concern when considering the generalizability of the FAW in conjunction with the inherent variability of patient populations. The best outcome for all patients could not be guaranteed, highlighting the need for staff education and new QC tools capable of testing these automated platforms.

Statistical

Statistical Process Control (SPC) trends QC results over time and can predict when a process starts to go out of tolerance before failure occurs. In a report of the application of SPC within radiotherapy, Xiao and Li⁵ stated the need to move away from conventional techniques that make binary decisions on predetermined tolerances, ignoring variability with the tolerances which does not support process improvement or mitigate against equipment failure and unplanned maintenance/repairs. Instead, move toward SPC for QC such as MLC and PSQC to be able to flag deviations from normal performance which could indicate a fault. Raveendran et al³⁵ states ignoring trends within conventional tolerances could lead to passive negligence and propose SPC to minimize variation, with the caveat that this is not always achievable in a clinical situation given human involvement. Raveendran et al³⁵ recommends using SPC to get machine, site and technique specific tolerances and actions for better PSQC monitoring.

Ma et al⁶³ performed SPC via the 6-sigma methodology⁶⁴ to determine MPC tolerances, as they questioned the independence of vendor tolerances. Two years of MPC results were collated to determine capability indices and associated tolerances and tested with 5 months of data. MPC vendor tolerances for isocentre size, MLC maximum offset, and rotation-induced couch shift were found to be too small. This knowledge reduced unnecessary engineer call outs and helped prevent fault blindness. All other calculated SPC tolerances were smaller than the vendor’s data and could identify a fault that resulted in machine breakdown that went previously unnoticed.

Able et al⁶⁵ proposed a model for preemptive maintenance of linacs by performing SPC on daily QC log files. Baseline tolerances were obtained by performing SPC on the log files of 7 linacs from 5 centers. Future log files were recorded and displayed on a dashboard against the tolerances. The dashboard was able to identify faults prior to occurrence and prompt preemptive maintenance. The authors stated that while it showed promise long-term monitoring, further evaluation will be required to establish the effectiveness of the model. Kawahara et al⁵⁸ performed FMEA to assess the biological effects to patients due to treatment interruption due to machine breakdown and stated that the method proposed by Able et al⁶⁵ would aid in reducing this risk.

Artificial intelligence

Research into the use of AI in radiotherapy is rapidly developing^{66,  67} and automated QC is a promising application of machine learning (ML).⁶⁸ A common theme of investigation was the use of AI models to predict the outcome of PSQC results.^69–73 Ono et al¹² found that, as of December 2023, more articles had been published on the use of AI within PSQC than in machine QC. AI has been trained on existing plans to link plan complexity metrics to PSQC results,^{70,  71} identify PSQC dosimetric errors,⁷⁴ identify planning data errors,⁷⁵ and predict actual MLC positions.^{51–53,  76} This information can be fed back into plan optimization^{52,  70} and/or can be used to reduce PSQC measurements.^{51,  71,  72} Concerns that too much confidence would be placed on the model and limit safe implementation exist.^{72,  73} Therefore, some researchers went a step further and calculated the uncertainty of the predictions to aid the user in decision-making.^{72,  73}

“Big data” of past QC results could be used to optimize future performance; by offering the opportunity to personalize QC pathways for patients with similar cases⁷⁷ or predicting machine faults.⁷⁸ Kalet et al⁷⁹ also queries whether vendors could use the linac log files they collect to develop specific maintenance and repair schedules. Ma et al⁸⁰ used daily QC results to develop a neural network model that could predict the record and trend of data variations and was robust when applied to a different radiotherapy machine and could foresee possible machine failure.

Mehrens and Kry⁸¹ used previous data obtained from Imaging and Radiation Oncology Core (IROC) thoracic audits and ML models to predict the pass/fail outcomes of dosimetry audits. Different treatment and motion-management approaches were found to influence audit performance. The results could be used to give centers personalized QC programs tailored to specific treatment setups.

E2E and patient-specific QC

An E2E technique examines a workflow and can identify potential failure modes between steps that would be missed by testing components in isolation. For example, Mertens et al⁸² considered the geometric uncertainty, which is increasingly important with the development of high precision treatments, in addition to the dosimetric uncertainty. Stating that it is hard to carry out these tests separately due to introduction of setup errors. Therefore, they developed an in-house 3D printed phantom that combined the 2. The use of phantoms to facilitate E2E is often employed, and literature pertaining to the design of novel phantoms for this purpose is common.^83–88 Ranging from phantoms designed to interrogate the entire workflow from CT/MR to patient treatment⁸⁶ to those designed for smaller sections of the pathway, such as SGRT only.⁸⁷

E2E testing is recommended in guidance, as part of commissioning or an annual QC²⁶ and as a method of testing a black-box system.¹⁹ Results of a 2024 UK linac QC survey³⁴ suggest mixed uptake with 49% of centers stating that E2E testing is performed, of which 21% clarified only as part of commissioning or external audits, and only 7% stated routine E2E QC (5% annually, 2% monthly). Loughery et al⁸⁹ produced a protocol for monthly E2E QC designed to detect complex errors not caught by conventional component-based QC. The protocol was validated by 3 centers and results were found to be consistent, suggesting that the protocol has potential use in future E2E QC initiatives. Schreiner¹¹ also advocated for the increased use of E2E within radiotherapy in a comprehensive review article. They state that while conventional QC verifies that individual steps are technically achievable, it does not consider whether the staff are carrying out the process correctly (human factors). They suggest this is where a regular E2E program could be a beneficial training and monitoring exercise for the performance of the entire system. Schreiner¹¹ caveats that while E2E testing catches errors conventional QC does not, E2E cannot be used in isolation as the testing is not specific enough to identify the cause of potential failures. In which case conventional QC records should be checked to troubleshoot.¹¹

Schreiner¹¹ presents 4 levels of QC tests, higher levels represent more system level testing: from level 1 component testing, to level 4 E2E of whole process. PSQC is defined as a level 3 QC test and thought to be a pseudo-E2E test. Despite this, Schreiner¹¹ states that PSQC should not be carried out in place of E2E tests, as audits have found the ability to identify errors as a result of PSQC is not always sufficient.^{90,  91} For example, Kry et al⁹⁰ compared in-house IMRT QC results with IROC QC results completed prior to clinical trials and found that in-house IMRT had a poor ability to detect a failing result. More recently, the SEAFARER study by Lehmann et al⁹¹ found centers passing their own PSQC, yet failing external audits which implied that centers could not detect clinically meaningful delivery errors. These results, alongside QC survey results, which stated PSQC was a significant burden on departments⁹² and plans were rarely modified as a result of PSQC results,^{34,  93} prompted Barnes¹³ to suggest a departure from measurement-based PSQC, provided a comprehensive sub-system QC program is in place.

The desire to reduce measurement-based PSQC is common. O’Daniel et al²³ used FMEA to investigate combining PSQC and conventional QC methods to determine the optimal combination for detecting errors that did not require measuring every plan. TG-100⁵⁴ was used to identify the failure modes that could be detected with PSQC. First, no PSQC was assumed which gave a detectability score of 0, and this was repeated for different combinations of PSQC types. Overall, a combination of log file-based QC with 3D dose calculation and MLC QC was determined to be optimal and allowed centers to consider a reduction in their measured QC. In 2025, O’Daniel et al⁹⁴ queried which errors PSQC should catch as part of a joint AAPM-ESTRO TG-360 subgroup. Failure modes that could cause greatest patient harm were identified and given priority for PSQC testing. To identify errors that machine QC alone would not detect, multiple methods of PSQC were required, from secondary dose calculation, transfer checks, and in vivo dosimetry.

Conversely, Mills and Colligan,⁶ proposed that the introduction of PSQC allows for a reduction of prospective QC (PQC). They argued that PQC demonstrates a machine is capable of treatment delivery, but the design of these prospective tests becomes too complex with the introduction of more advanced delivery techniques. As a result, PSQC becomes unavoidable as it verifies treatment delivery. However, given limited resources, it is impractical to increase PSQC without the reduction of PQC. Therefore, Mills and Colligan⁶ proposed a framework that allows for the reduction of PQC by approximately 50%, provided that weekly PSQC is performed on each linac.

Audits and clinical trials

Dosimetry audits are facilitated by national and international audit groups, for example, IPEM interdepartmental dosimetry audit group (IDA) in the UK,⁹⁵ IROC in the US,⁹⁶ BELdART in Belgium,⁹⁷ ACDS in Australia,⁹⁶ and the IAEA internationally.⁹⁸ Dosimetry audit credentialing is often a requirement for participation in clinical trials to ensure that a center is capable of specific clinical delivery techniques and to reduce variable center-specific practice from influencing results, and potentially underpowering clinical trials. Again, these are facilitated by approved groups, for example, RTTQA in the UK,⁹⁵ JCOG in Japan,⁹⁹ and DMCG in Denmark.¹⁰⁰

There have been many review articles on dosimetry audits. Clark et al²⁵ published a review article in 2015, concluding undertaking regular audits help modernize practice and ensure prescribed dose is delivered. Ibbott and Thwaites¹⁰¹ published a review regarding dosimetry audits for more complex techniques. Muhammad et al¹⁰² wanted to ensure that for accuracy in radiation dose delivery, external audits had an equal importance with routine in-house QC. Pasler et al¹⁰³ wrote a review article on novel methodologies for dosimetry audits in 2018, which included the use of deformable and anthropomorphic phantoms, as well as novel approaches to non-phantoms such as log files and EPID measurements, but stated that the harmonization of dosimetry audit techniques are required first. Brooks et al¹⁰⁴ focused on auditing outlining in clinical trials, while not machine QC, outlining is still an important part to audit if considering the full radiotherapy pathway.

Lehmann et al⁹¹ proposed a new concept for validating PSQC for clinical trial participation. Centers sent treatment plans to the audit team. Delivery errors were introduced to the plans and returned. Participating centers performed PSQC on these plans and informed the audit team of their results. Centers were not always able to detect the clinically significant delivery errors and often passed plans that increased the risk of spinal cord injury. Also, sometimes plans were failed that had no clinically significant impact. No consistency was found between centers, leading Lehmann et al⁹¹ to state that the sensitivity of PSQC should be audited, optimized, and justified as clinical trial data is pooled and therefore consistency is critical. On a similar theme, Brooks et al¹⁰⁵ developed reference plans with introduced errors to compare various audit methodologies within the Global Harmonization Group (GHG),¹⁰⁶ as part of the GHG work to test the comparability of audit systems and align international audits across the globe. Jomy et al¹⁰⁷ performed a systematic review between 2021 and 2024, which showed that deviations from trial protocols were linked to worse clinical outcomes. Retrospective QC was not shown to correct existing issues. Deviations were found in contouring, planning, and dose delivery, proving the importance of pretrial participation E2E audits.

Looking to the future, Misson et al⁹⁵ published a position paper on audits in the UK. IPEM has interdepartmental audit (IDA) groups setup across the UK in which regions work among themselves to carry out annual dosimetry audits. NPL and RTTQA may perform voluntary or obligatory audits prior to centers going clinical with new techniques and equipment and if they request to participate in clinical trials. Misson et al⁹⁵ proposes to centralize these audits, to allow IDA groups to benchmark against each other, and reduce demand on NPL and RTTQA.

Nielsen et al¹⁰⁰ gives recommendations for a comprehensive QC program and states a robust framework is essential for safety. They suggested that FMEA could be integrated into the clinical trial QC design which would allow QC protocols to be customized to the study, resources to be directed to QC failure modes that have the highest potential impact on trial outcomes, and identify processes that are most variable among centers to provide targeted support.

Kerns et al¹⁰⁸ extended the value of dosimetry audits using E2E tests by producing an additional test for IROC thorax phantom audits. They created a beam model to facilitate independent recalculation of centers plans. A comparison of the center versus IROC plan calculation indicated if failing measurements were a result of calculation errors in the treatment planning system.