Evaluation Framework for Deep Neural Networks in Lymphoma Segmentation from PET/CT
Automated lymphoma segmentation from PET/CT images faces critical gaps, including a lack of out-of-distribution testing and limited comparison with expert annotators. This study proposes a clinically relevant evaluation framework for deep segmentation networks, emphasizing lesion-specific measures alongside overall accuracy. The framework accounts for intra- and inter-observer variability to better assess task difficulty. It was demonstrated on four deep networks: ResUNet, SegResNet, DynUNet, and an unnamed fourth model. The study argues that precise lesion detection and disease quantification require such measure-specific evaluations. The paper, available on arXiv with identifier 2311.09614v5, highlights the need for robust, generalizable models in medical imaging.
Key facts
- Study addresses gaps in automated lymphoma segmentation from PET/CT images.
- Out-of-distribution testing is rarely incorporated in existing literature.
- Model performance should be compared with expert human annotators.
- Intra- and inter-observer variability are considered to understand task difficulty.
- Most approaches overlook lesion-specific measures.
- A clinically relevant evaluation framework is proposed.
- Four deep networks are assessed: ResUNet, SegResNet, DynUNet, and a fourth.
- The paper is arXiv:2311.09614v5, announced as replace-cross.
Entities
—