Room: AAPM ePoster Library
Purpose: being the standard metric in IMRT QA, gamma analysis has two shortcomings: (1) it lacks sensitivity to some critical errors (2) it does not provide effective means to classify the source of errors. We aim to achieve error detection and classification simultaneously with dual neural networks (DNN): a convolutional neural network (CNN) analyzing a distance-to-agreement (DTA) map for spatial errors and a general artificial neural network (ANN) analyzing a dose difference (DD) histogram for dosimetric errors.
Methods: a pair of measured and calculated 2D dose distributions, we extracted (1) DTA maps along x and y directions and (2) a histogram of relative DD. We designed a 4-layer CNN to analyze DTA maps and a 3-layer ANN to analyze the histogram. The CNN classifies seven categories of spatial errors: incorrect effective source size, MLC leaf bank errors (±1 mm), single MLC leaf positioning errors (±2 mm), and device misalignment errors (1 mm in x or y direction). The ANN classifies six categories of dosimetric errors: incorrect MLC transmission (±1%), and four types of MU scaling errors (±1% & ±2%). An in-house planar dose calculation software was used to simulate measurements with and without introduced errors. Both networks were trained and validated with 13 IMRT plans for head and neck, lung, and prostate treatments. A 5-fold cross validation technique was used to evaluate their accuracy.
Results: a PC equipped with a single GPU, the training of the CNN and ANN completed in 89 and 10 seconds, respectively. The average classification accuracy across the five folds was 95.6±1.5% and 98.3±0.7% for the CNN and ANN, respectively.
Conclusion: to gamma analysis, the proposed DNN approach analyzes DTA and DD separately, enabling simultaneous error detection and classification with high accuracy. Complementary to gamma analysis, it could potentially shift the paradigm of IMRT QA.