Program Information
Clustering Breathing Curves in 4D Radiotherapy Using Two Machine Learning Algorithms
Q Li1*, M Chan2 , B Wang3 , C Shi2 , (1) The Graduate Center of City University of New York, New York, New York,(2) Memorial Sloan-Kettering Cancer Center, Basking Ridge, NJ, (3) Univ Louisville, Louisville, KY
Presentations
TU-C3-GePD-J(A)-2 (Tuesday, August 1, 2017) 10:30 AM - 11:00 AM Room: Joint Imaging-Therapy ePoster Lounge - A
Purpose: Respiratory motion is an existing clinical challenge for cancer treatment. Besides the current approaches, it is important to investigate a novel method to cluster the patient breathing curves. This study is to predict the tumor motion using machine learning (ML) techniques.
Methods: Two ML clustering algorithms: k-means and hierarchical were used to analyze 341 breathing curves obtained from 4DCT. After cleaning the data (i.e., correcting the baseline drifting, interpolating the missing data), 74 high-quality datasets were retained for further analysis. A visual examination was performed on the breathing signals and its frequency spectrum. Other exploratory research was performed (i.e., correlation analysis, statistical distribution analysis) to examine the violation of model assumptions. For validation, a small sample data set (10 breathing curves) with obvious signal behaviors was selected: one group with regular breathing (8 curves in total, among them 1 curve with nice and slow breathing) and the other with irregular breathing (2 curves). Six features (average amplitude, standard deviation of amplitude, dominant frequency, maximum power, standard distribution of power, secondary frequency power) were extracted from the time series and frequency spectrum.
Results: Both algorithms completed the task successfully with 100% accuracy distinguishing irregular breathers from regular breathers. By specifying 3 clusters in k-means, the perfect breather can be distinguished from the regular breather group. K-means is extremely sensitive to cluster center initialization, and the results may not be reproducible and may lack consistency. Hierarchical clustering could generate more consistent results; however, it requires a quadratic amount of time used in k-means, which could lead to inefficiency for large data sets.
Conclusion: Our results show that it is possible to cluster patients' breathing patterns using ML tools. However, further study is needed to correlate the results with other clinical information of patients before they can fully benefit at clinical application.
Contact Email: