Click here to

Session: Patient Specific Quality Assurance [Return to Session]

Variational Clustering: Toward Unsupervised Machine Learning-Based Error Detection in Patient-Specific IMRT QA

J Perez-sanchez*, B Lu, S Samant, J Li, C Liu, G Yan, University of Florida, Gainesville, FL


MO-E115-IePD-F4-4 (Monday, 7/11/2022) 1:15 PM - 1:45 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 4

Purpose: Deep learning techniques have been used to enhance error detection in IMRT QA. These techniques rely on supervised learning, meaning the model can only recognize errors presented in the training. The purpose of this work is to investigate the feasibility of using unsupervised variational clustering technique for error detection in IMRT QA.

Methods: A Variational Autoencoder (VAE) consisting of an encoder, a decoder, and a loss function was used to achieve variational clustering. It probabilistically encodes the input to latent distributions. Inputs sharing similar features are closely encoded, enabling automatic clustering. Here we show that a VAE could be used to cluster IMRT QA data based on error categories. Dose distributions from 20 IMRT plans (122 fields) were collected. Three error types were simulated (1 mm MLC leaf bank shift, 2 mm single leaf shift, and 1 mm phantom shift). Distance-to-agreement maps were calculated to create 488 images as input. The VAE had a 2-convolution-layer encoder, a 2-dimensional latent space, and a 3-transpose-convolution-layer decoder. The loss function included the reconstruction loss and the Kullback-Leibler divergence. No labels were provided in the training (unsupervised). A test set of 48 images were used to test the classification accuracy. Test images were input to the trained encoder and classified by their latent distribution.

Results: The maximum likelihood estimates for the underlying distribution were used to compute the posterior probability of a point belonging to each cluster. Only one input image was misclassified (a phantom shift error was misclassified as MLC bank shift error). The overall classification accuracy was 97.9%.

Conclusion: An unsupervised variational clustering technique implemented with a VAE was demonstrated for error detection in patient-specific IMRT QA. It has the potential for automatically clustering input based on error types in the data with high accuracy.


Quality Assurance, Radiation Therapy, Feature Extraction


TH- External Beam- Photons: Quality Assurance - IMRT/VMAT

Contact Email