Room: AAPM ePoster Library
Purpose: propose a deep learning model for improved segmentation accuracy in PET/CT images. Automated tumor segmentation from Fluorodeoxyglucose (18FDG)-positron emission tomography and computed tomography (18FDG-PET/CT) has the potential to aid diagnosis and treatment response evaluation in multiple tumor systems. Manual delineation is time consuming and subject to human bias. Conventional segmentation methods are often obscured by radiotracer uptake from normal tissues, including brain, heart, liver, bilateral kidneys, and bladder. Specifically, heart, liver, and kidneys are variably avid on routine 18FDG-PET examinations, imposing additional challenge for accurate segmentation.
Methods: this work, we propose a semantic segmentation network based on autoencoder architecture reinforced by an object detection branch for segmenting normal tissue and tumor from 3D 18FDG-PET/CT. The autoencoder network consists of symmetrical encoder and decoder with ResNet blocks and skip connections. The object detection branch is derived from You only look once (YOLO) -real time object detection which is added at the end of the encoder. Our hypothesis is that the object detection branch not only regularizes the network but also imposes location constraints based on anatomy.
Results: test our hypothesis in a limited cohort consists of 34 pediatric patients with Hodgkin lymphoma with pre treatment PET/CT images. We used F1 score and Hausdorff distance at 95th percentile to quantify segmentation accuracy and compare the results with established networks. The proposed method achieved the overall best segmentation accuracy and improved the average F1 score by 0.19 in normal tissues and tumor segmentation compared to V-Net. The overall Hausdorff distance improved by 95% compared to V-Net.
Conclusion: Automatic detection and segmentation of normal tissues and tumors by combining segmentation and regularization networks substantially improves the clinical detection and treatment response evaluation of PET and CT images relative to existing models.