Using Fourier Analysis and Machine Learning to Detect Heart Irradiation In Deep Inspiration Breath Hold Breast Radiotherapy
Abstract
Purpose
Develop an automated method of detecting heart irradiation in Deep Inspiration Breath Hold (DIBH) breast cancer radiation therapy (RT).
Methods
EPID cine frames were exported for all tangential beams of 10 DIBH breast RT treatments. For the cine series of each tangential beam, Fourier analysis was performed using the time series of each image pixel. Frequency-domain images were generated for each tangential beam, identifying the frequency fluctuations of each pixel. The heart's periodic motion in the tangential field highlights the heart edge. The frequency-domain image with the highest peak Signal-to-Noise Ratio (SNR) from each tangential beam was selected, and the corresponding plot of SNR per row was used to train supervised machine learning to distinguish cases with and without heart irradiation. EPID cine frames from 398 tangential beams were used in a five-fold cross validation, with 221 beams featuring heart irradiation. The machine learning models included Naïve Bayes, K-Nearest Neighbours (KNN), Support Vector Machine (SVM) and Logistic Regression. The area under the Receiver Operating Characteristic (ROC) curve was used to evaluate each model's performance.
Results
The highest area under the ROC curve was 0.81 (KNN), while the lowest was 0.60 (SVM). Errors in classification were associated with anatomical structures, such as the chest wall, mimicking heartbeat motion in the frequency-domain images. Variations in pixel brightness within the lung along the superior/inferior direction also contributed to misclassification. Differences in signal strength of the heart edge in frequency-domain images were related to variations in contrast between the heart and lung across different EPID cine series.
Conclusion
Fourier analysis combined with supervised machine learning was used to automate the detection of tangential heart irradiation in DIBH breast RT. Future work will explore direct analysis of EPID images using a Convolutional Neural Network (CNN) to identify heartbeat motion with improved robustness.