Author: Ling Chen, Alexei V. Trofimov, Yi Wang, Dufan Wu 👨🔬
Affiliation: Massachusetts General Hospital, MGH 🌍
Purpose:
Selecting gaze angles of the eye is an important part of set-up of proton therapy for ocular tumors, ensuring that the treatment beam could irradiate the tumor while maximally sparing important healthy structures. This study aims to develop machine learning algorithms to select appropriate gaze angles for planning CT scans, depending on the target size and location within the eye.
Methods:
A dataset comprising 1,477 treatment plans for ocular tumors was collected for this study. Each plan includes six input features: tumor’s "clock hour" position in the eye, tumor dimensions (length, width, height), distances from the tumor to optic disc and macula. The outputs are two variables selected by expert planners: the polar and azimuthal angles of the gaze direction. We divided the two outputs into 11 and 12 classes, respectively, to frame the problem as a classification task. Among the collected data, 1,329 treatment plans were used for training, while the remaining 148 plans were reserved for testing. Four machine learning-based classification algorithms were evaluated in this experiment: K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP) and Random Forest (RF). The 10-fold cross-validation method was used to objectively measure the performance of the methods.
Results:
The prediction accuracies for the polar angle at 2° precision using KNN, SVM, MLP, and RF were 36.8%, 38.0%, 36.9%, and 39.2%, respectively, while for the azimuthal angle at 30° precision, they were 38.0%, 39.3%, 43.5%, and 43.7%. At 6° precision, the polar angle accuracies were 70.9%, 67.0%, 68.9%, and 70.9%, and for the azimuthal angle at 90° precision, the accuracies were 68.9%, 74.9%, 75.9%, and 76.6%. Overall, RF achieved the best performance.
Conclusion:
Given the six input variables, random forest could achieve highest accuracy in predicting the gaze angles with a precision of 6° in polar and 90° in azimuth.