1096 - Automated Segmentation of Pancreatic Tumor Microvasculature in High-Resolution Optical Coherence Tomography (OCT) Using Pre-Trained Large-Scale Supervised Networks for Early Evaluation of Radiation Th

E. Abouei¹, Y. Li¹, J. Zabel², H. Contreras², A. Vitkin³, and X. Yang¹; ¹Emory University, Atlanta, GA, ²University of Toronto, Toronto, ON, Canada, ³Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada

Purpose/Objective(s): Speckle variance OCT (svOCT) can visualize tissue microvasculature in vivo in 3D, thereby offering insights into the tumor microenvironment and enabling longitudinal monitoring of vascular response to radiotherapy (RT). To investigate pancreatic tumor response to both conventional RT and stereotactic body RT (SBRT), our study focuses on the development of an automatic segmentation framework with the objective of enabling quantitative assessment of tumor microvascular network properties, such as vessel lengths and tortuosity. Materials/

Methods: We propose a deep learning framework for automatic vascular segmentation in pancreatic tumor. For our algorithm, we adopted an encoder-decoder network which is well-suited to spatially aggregate feature maps across several scales. We developed a large-kernel convolutional neural network, utilizing the scalable and transferable architecture. In the encoder of our network, we replaced the traditional 3x3x3 convolution kernels with 13x13x13 kernels, which captures larger spatial contexts vital for accurate segmentation. We also added squeeze-and-excitation modules to facilitate feature recalibrations. 3D svOCT images were obtained over a 6x6 mm² area with 800 A-scans per frame and an imaging depth of 1.5 mm in tissue, over a period of 7 weeks at various time points following RT. OCT images were acquired with lateral and axial resolutions of 15 and 8 µm in air, respectively. Due to the time-intensive nature of obtaining semi-manual segmentations in Matlab, we only employed a small dataset consisting of 10 svOCT images for the fine-tuning and testing of our proposed model. These semi-manual segmentations served as ground truth for quantifying the accuracy of our automatic deep learning approach, evaluated using the Dice score.
Results: The Dice scores ranged from 0.74 to 0.84 with an average value of 0.79. The relatively modest scores likely stem from the limited size of the dataset and challenges in accurately defining vessel boundaries inherent in svOCT images. Since the svOCT studies focus on relative changes in the derived microvascular metrics between control and experimental groups, and/or within the same group over time rather than the absolute values, the observed lower scores will likely suffice for practical applications.
Conclusion: Using a deep learning model to segment high-resolution OCT imaging provides a promising approach for automated analysis of tumor microvasculature, potentially facilitating the understanding of tumor progression and response to RT. Our next endeavor involves leveraging a comprehensive svOCT dataset to refine our model and enhance segmentation accuracy. Then we plan to apply our segmentation method for quantitative analysis of control and irradiated mice to investigate tumor response to conventional RT and SBRT.