MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Study On Breathing Pattern Classification and Prediction Using Machine Learning Algorithms

X Tang1*, Y Ou2 , Z Saleh3 , J Jeong4 , W Cai5 , Y Song6 , M Zhang7 , M Chan8 , C Shi9 , (1,3,4,5) Memorial Sloan Kettering Cancer Center, West Harrison, NY, (2) Columbia University, New York, NY, (6, 7) Memorial Sloan Kettering Cancer Center, Montvale, New Jersey, (8,9) Memorial Sloan-Kettering Cancer Center, Basking Ridge, NJ.

Presentations

(Sunday, 7/29/2018) 4:00 PM - 4:55 PM

Room: Room 202

Purpose: Managing the breathing motion related moving tumors is challenging and requires knowledge of real-time tumor location. We herein analyze and classify breathing patterns, which is an important step towards predicting tumor location in real-time. Machine learning based models will be built to predict which breathing pattern a given new breathing curve belongs to.

Methods: 1774 breathing curves acquired by the RPM system were included. First, the curves were classified. One supervised learning algorithm (artificial filters based on quantiles and standard deviations) and two unsupervised learning algorithms (k-means and hierarchical) were applied to classify the data. For each classification result, we tested the following prediction models: k-nearest neighbors, random forest, linear discriminate analysis and neural network. Handout validation was used to evaluate the prediction accuracy of each method. The silhouette information was also computed. Visualization of the classification methods with plots was analyzed.

Results: In general, the unsupervised algorithms outperformed the supervised algorithm in terms of prediction accuracy. The hierarchical clustering and neural network prediction combination had the highest accuracy of 99.43%. The silhouette of hierarchical clustering shows that cluster 1 and 3 contains most of the units. Cluster 3 and cluster 6 have relatively high average silhouette widths. However, their cluster sizes might be too small to the meaningful for training models. For artificial clustering visualization, there were many overlapping areas among different clusters. For k-mean clustering, 5 clusters contain majority of the data. The hierarchical clustering visualization demonstrated that it seemed not being able to differentiate breathing patterns with different amplitudes.

Conclusion: We have successfully classified the breathing curves using supervised and unsupervised algorithms. The supervised method tended to have many overlapping areas. The unsupervised learning algorithms resulted higher prediction accuracy and seemed better candidates for breathing curve classification.

Keywords

Gating, Motion Artifacts, Organ Motion

Taxonomy

TH- External beam- photons: Motion management (intrafraction)

Contact Email