Encrypted login | home

Program Information

Deep Nets Vs Human Designed Features in Medical Physics: An IMRT QA Case Study

no image available
G Valdes

V Rideout , Y Interian , T Solberg 2 , G Valdes 2,*, 1.MS in Analytic Program, University of San Francisco. 2. UCSF Comprehensive Cancer Center, San Francisco, CA.


TU-L-GePD-J(B)-5 (Tuesday, August 1, 2017) 1:15 PM - 1:45 PM Room: Joint Imaging-Therapy ePoster Lounge - B

Purpose: To compare the performance of Deep Neural Nets developed without domain knowledge against carefully designed features by domain experts to predict gamma passing rates for IMRT QA.

Methods: 498 IMRT plans delivered on TrueBeam and Clinac IX linacs (Varian Medical Systems) were measured with MapCHECK2 (Sun Nuclear Corp) using gamma criteria of 3%local/3mm with a 10% threshold. A fluence map was calculated for each plan using a Matlab script (Matlab Inc). This fluence maps were used as input for Convolution Neural Nets (CNN). CNNs were trained using the Keras package and Theano Python library. An architecture inspired by the convolutional blocks in the VGG16 ImageNet winning model was used. Gaussian and Poisson distributions were used in the output layer. Synthetic data was created to boost the performance of the CNNs by rotating and translating the fluence maps during training. Dropout with p=0.5 was used for regularization between the hidden dense layer and the output layer . For the supervised Machine Learning algorithm, a previously published generalized linear model (GLMNET), specifically Poisson regression, with 89 human designed features that describe different failure modes between planning (Eclipse, Varian Medical System) and linac delivery was used.

Results: If no information about energy or Linac is provided CNN and GLMNET have similar performance. However, if information about the energy of the plan is provided, GLMNET outperforms the CNNs in term of mean value and number of outliers.

Conclusion: Although Deep Nets can learn features by themselves and predict IMRT QA results without expert intervention, they were outperformed by human designed features. This is likely to be the case in datasets of similar size, regardless of the problem.

Contact Email: