Abstract

Cecile Wolfs¹, Renate van Doormaal¹, Evelyn de Jong², Richard Canters¹, Frank Verhaegen¹

¹GROW School for Oncology, Maastricht University Medical Centre+, Department of Radiation Oncology (Maastro), Maastricht, The Netherlands; ²Catharina Hospital, Department of Radiotherapy, Maastricht, The Netherlands

Electronic portal imaging device (EPID) dosimetry aims to detect errors before or during radiotherapy treatment. Clinically, threshold classification methods (e.g., 10% gamma fail rate threshold) are used for detecting errors, but these methods lead to loss of information (compressing multi-dimensional EPID data into a few numbers) and cannot be used for identifying causes of errors. Advanced classification methods, such as deep learning, can use the available multi-dimensional information as input. In this study, convolutional neural networks (CNNs) were trained to detect and identify error type and magnitude of simulated treatment errors in lung cancer patients. The purpose of this simulation study is to show the applicability of CNNs for error identification, using EPID dosimetry in various scenarios.

Three lung cancer patient cohorts and various EPID dosimetry modalities were considered: (1) 69 stereotactic body radiotherapy (SBRT) plans with 2D time-integrated (TI) pre-treatment EPID dosimetry, (2) 47 VMAT or hybrid plans (with both static beams and VMAT arcs) with 2D-TI transit EPID dosimetry and (3) 46 VMAT plans with 2D time-resolved (TR) transit EPID dosimetry. Clinically realistic ranges of mechanical errors (all cohorts), anatomical changes (cohort 2 and 3) and positioning errors (cohort 2 and 3) were simulated (Figure 1). Predicted portal dose images (PDIs) containing errors were compared to error-free PDIs using (3%, 3mm) gamma analysis. For cohort 1, (3%, 1mm) gamma analysis was additionally performed and evaluated. CNNs were optimized and trained to classify errors using gamma maps as input. Three classification levels were assessed (Figure 1), from coarser to more detailed: Level 1 (main error type, e.g., anatomical change or mechanical error – cohorts 2 and 3), Level 2 (error subtype, e.g., tumor regression or systematic MLC shift – all cohorts) and Level 3 (error magnitude, e.g., >50% tumor regression or MLC shift > 1 mm – all cohorts).

CNNs generally showed good performance for all classification levels (Table 1), with overfitting becoming slightly more apparent for the more detailed classification levels. For cohort 1, (3%, 1mm) gamma maps provided consistently better results than commonly used (3%, 3mm) gamma maps. As the CNNs used for cohort 3 were optimized for TI data, but not yet for TR data, their performance is lower.

Deep learning is a promising powerful tool for identifying types and magnitudes of treatment errors with EPID dosimetry. It can advance the field of EPID dosimetry substantially, by providing rapid, automated models for error identification, and with that, additional detailed information not currently available from EPID dosimetry. Currently, the applicability of these CNNs to clinically measured EPID data, as well as optimization of CNNs for TR data and the use of different gamma criteria and dose comparison methods as model input are under further investigation for these lung cancer cohorts.