Application of a Convolutional Neural Network to the Segmentation of Lungs in Mice
E Criscuolo1*, R Sali2, E Graves3, L Soto4, (1) University of Connecticut, Storrs, CT, (2) Stanford University, Stanford, CA ,(3) Stanford University, Stanford, CA, (4) Stanford University, Stanford, CA,
Presentations
PO-GePV-M-212 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: In preclinical radiotherapy using small animals, segmentation of targets for irradiation is critical to ensure consistency of treatments and interpretability of results, however commonly this procedure is done manually in a user-dependent fashion. Convolutional neural networks (CNNs), with their high-level feature detection, offer a promising alternative. A well-trained algorithm has the potential to rapidly and consistently segment areas of interest. With this in mind, the purpose of this study was to develop a CNN capable of segmenting lungs within computed tomography (CT) scans of mice as a proof of principle for ML-based target definition for preclinical radiotherapy.
Methods: A dataset of 30 thoracic microCT scans was obtained from a previous experiment involving single lung irradiation. The RT_Image software package and a region growing algorithm were utilized to define regions of interest (ROIs) covering the lungs to serve as a training dataset for the neural network. After processing, these images were divided in a 20:5:5 ratio for training, validation, and testing respectively. The CNN model UNET (University of Freiburg) was trained with 100 epochs of training, a rotation range of 90 degrees for image augmentation, and the Adam optimizer with default parameters.
Results: Preliminary results of the trained network demonstrated an ability to segment lung volumes, however with considerable room for improvement. We are currently pursuing strategies to provide more faithful lung segmentation including additional datasets, more rigorous manual segmentation in the training set, and additional data augmentation. Our results indicate that a more comprehensive training dataset will lead to a high accuracy volumetric segmentation.
Conclusion: A CNN approach to automating lung definition for preclinical radiotherapy shows promise to improve the practice of small animal radiation studies. With this ability more soundly developed, the trained network will help contribute to the standardization of practices in preclinical small-animal research.
Funding Support, Disclosures, and Conflict of Interest: Funding was provided through the AAPM Undergraduate Research Fellowship.
Keywords
Not Applicable / None Entered.
Taxonomy
IM/TH- Image Segmentation Techniques: Machine Learning
Contact Email