ESTRO 2024 - Abstract Book

S3182

Physics - Detectors, dose measurement and phantoms

ESTRO 2024

produce a full set of data to compare against baseline in under one hour, offering a new approach to small-field dosimetry.

Material/Methods:

A 2mm thick polystyrene scintillating sheet is sandwiched between 2 non-scintillating blocks and contained within a light-tight commercial case along with an SLR camera. This self-contained unit provides a simple fixed geometry setup and when placed on the treatment couch is ready for data acquisition. The case is irradiated with a vertical, fixed source to surface distance 6MV (flattening filter free) beam generated by a CyberKnife linear accelerator using fixed circular collimator with diameters ranging from 5mm to 60mm. The resulting images from each beam diameter are analysed to generate output factors at a depth of 15mm for each collimator compared to the 60mm reference beam. The image from each collimator exposure can also be used to generate a percentage-depth-dose (PDD) curve within the block, and to extract the beam profile at any desired depth. A convolutional neural network was trained on Monte Carlo simulated images of the experimental setup in order to predict the dose given on the measured photograph. Validation of the model was performed using an independent dataset. The final data from these processed images was compared to classically obtained gold-standard data acquired using a microdiamond detector within a watertank.

Results:

Classical image processing of the resultant images produced PDD, profile and output factor data that was in reasonable agreement with the physical chamber results, but not of sufficient accuracy for radiotherapy quality assurance. Once the images were processed using the central neural network, the mean absolute difference between the processed image PDD and the measured dose with the microdiamond chamber was around 1%. Across all collimator widths from 5mm to 60mm, 95% of points met the gamma 3%/3mm criteria, and 78% met the 1% 1mm criteria. For output factors, the mean absolute difference between that derived from the processed image and that determined with a microdiamond detector was 1.1%.

Conclusion:

We have demonstrated that machine learning methods in radiotherapy dosimetry are able to offer quantitative predictions of out-put factor in small fields dosimetry.

We have produced a prototype of a device that could quickly set-up to measure and check small field radiotherapy data against a reference without the need of a watertank or ionisation chamber. From a single set of photographs an output factor curve, PDD and profile data could be obtained and acquired in less than one hour, with an accuracy equivalent to classical methods. We envisage this being used for regular, relative quality assurance to assess an accelerators performance, or to quickly validate machine data following a repair/service.

Keywords: scintillation, deep-learning