3302 - Convolutional Neural Network (CNN)-Based Segmentation Method for Automatic Clinical Target Volume (CTV) Contouring in Strut-Based Applicator Brachytherapy in Breast Cancer

12:45pm - 02:00pm PT

Hall F

Screen: 34

POSTER

Presenter(s)

Minh Nguyen, MS, BS - East Carolina University Brody School of Medicine, Greenville, NC

M. Nguyen¹, J. W. Jung², H. Kadji³, A. V. Hnatov⁴, and A. W. Ju²; ¹Brody School of Medicine at East Carolina University, Greenville, NC, ²Department of Radiation Oncology, East Carolina University, Greenville, NC, ³Department of Radiation Oncology, University Hospitals Seidman Cancer Center, Cleveland, OH, ⁴ECU Health Radiation Oncology, Greenville, NC

Purpose/Objective(s): Accurate delineation of the Clinical Target Volume (CTV) is crucial for effective radiotherapy planning in breast cancer treatment. Manual contouring is time-consuming and prone to variability. We explore the use Convolutional Neural Network (CNN)-based segmentation method for automatic CTV contouring in Strut-Adjusted Volume Implant (SAVI)® brachytherapy, aiming to improve efficiency and consistency in radiotherapy planning.

Materials/Methods: CT images from 200 breast cancer patients post-SAVI catheter implantation were collected. A radiation oncologist manually contoured CTV (ground truth). 2D U-NET architecture was used for segmentation. Data was split 72% training, 18% validation, and 10% test sets. A post-processing threshold of 0.5 was applied to reduce false positives. Performance was evaluated using Dice-Sorensen Coefficient (DSC). Training sets (n=200) were resampled using a bootstrap technique to improve the model’s robustness.

Results: The best agreement between automatic and manual CTV was achieved when the dataset was split by left and right breast. Average DSC: left breast 0.86, right breast 0.82, combined 0.80. Right breast dataset showed more false positives, potentially due to anatomical differences. With 50 epochs, training took <7 hours. With the bootstrapping, the mean DSCs were 0.834 and 0.828 for the left and right breasts, respectively.

Conclusion: This study demonstrates the promising application of AI in streamlining and enhancing the radiotherapy planning process for SAVI brachytherapy in breast cancer patients. AI shows potential to reduce contouring time, improve consistency and accuracy, and assist physicians by capturing errors and improving workflow. While this model was developed for SAVI brachytherapy, future work will explore its applicability to other brachytherapy techniques and external beam radiotherapy planning.

Abstract 3302 - Table 1: Model performance with Bootstrapping (n=200)

	Dice Similarity Coefficient
Laterality	Mean	SD	Min	Max
Left Breast	0.834	0.067	0.509	0.937
Right Breast	0.828	0.069	0.521	0.933