(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: treating intact-breast cancer, tangential beams are manually placed to cover the whole breast. We investigate automating tangential beam placement for left-sided intact breast cancer utilizing fully-convolutional networks.
Methods: scans and treatment plans from 599 intact-breast patients treated with opposed tangential fields to the whole left breast were used in this study. Patients were split into training (n=359)/cross-validation (n=120)/test(n=120) sets. A 3D fully-convolutional neural network was trained to autosegment the 50% iso-dose volume taking into account the location of the tumor bed. This volume was then imported into TPS to guide tangent beam placement. A script with graphic user interface was created to guide users for the operation. The predicted and clinical beams are compared in terms of gantry angle, collimator angle and medial collimator size for a subset of test set (n=25).
Results: gantry angle difference between the clinical and predicted beams was 3.3 degrees (range: 1-8), average collimator angle difference was 2.3 degrees (range: 0-8), and average medial collimator size difference was 0.4 cm (range: 0-1.0). Patient disease presentation varied largely: one patient needed axilla coverage, one patient had significant medial extent of the tumor bed. For these two cases, differences between clinic and predicted values were relatively large, whereas for patients where disease was found more central to the breast the differences were found to be smaller.
Conclusion: automated tangential beam placement tool for left-side intact breast cancer was created using deep learning. The predicted beams’ angles/distances agreed well with clinical beams within a few degrees and millimeters for most cases.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
