AutoPET Challenge 2022: Automatic Segmentation of Whole-body Tumor Lesion Based on Deep Learning and FDG PET/CT
This work addresses the problem of systemic tumor segmentation for medical imaging analysis, which is incremental as it builds on existing deep learning methods for a specific domain challenge.
The paper tackled the challenging task of automatic whole-body tumor lesion segmentation from PET/CT scans by proposing a novel deep learning training strategy, achieving a Dice score of 0.7574, false positive volume of 0.0299, and false negative volume of 0.2538 on a preliminary test set.
Automatic segmentation of tumor lesions is a critical initial processing step for quantitative PET/CT analysis. However, numerous tumor lesion with different shapes, sizes, and uptake intensity may be distributed in different anatomical contexts throughout the body, and there is also significant uptake in healthy organs. Therefore, building a systemic PET/CT tumor lesion segmentation model is a challenging task. In this paper, we propose a novel training strategy to build deep learning models capable of systemic tumor segmentation. Our method is validated on the training set of the AutoPET 2022 Challenge. We achieved 0.7574 Dice score, 0.0299 false positive volume and 0.2538 false negative volume on preliminary test set.The code of our work is available on the following link: https://github.com/ZZZsn/MICCAI2022-autopet.