Clinical application of AI for lung lobe segmentation and functional quantification with SPECT/CT: open-source versus custom models.
Authors
Affiliations (12)
Affiliations (12)
- Department of Nuclear Medicine and Molecular Imaging, The Ottawa Hospital, Ottawa, ON, Canada. [email protected].
- Department of Systems and Computer Engineering, Carleton University, Ottawa, ON, Canada. [email protected].
- The Ottawa Hospital Research Institute, 1053 Carling Ave, P.O. Box 232, Ottawa, ON, K1Y 4E9, Canada. [email protected].
- Department of Cell and Systems Biology, University of Toronto, Toronto, ON, Canada.
- Department of Medicine, University of Ottawa, Ottawa, ON, Canada.
- Faculty of Medicine, University of Ottawa, Ottawa, ON, Canada.
- Department of Radiology, The Ottawa Hospital, Ottawa, ON, Canada.
- Faculty of Medicine and Health Sciences, McGill University, Montréal, Québec, Canada.
- Department of Radiology, Radiation Oncology and Nuclear Medicine, Faculty of Medicine, Université de Montréal, Montreal, Canada.
- Department of Nuclear Medicine and Molecular Imaging, The Ottawa Hospital, Ottawa, ON, Canada.
- Department of Systems and Computer Engineering, Carleton University, Ottawa, ON, Canada.
- Division of Nuclear Medicine and Molecular Imaging, Department of Medicine, University of Ottawa, Ottawa, ON, Canada.
Abstract
Lung lobe quantification with SPECT/CT is essential for treatment planning but remains clinically limited by inadequate tools. Although AI-based approaches show promise, their clinical validation is limited. This study compares the clinical accuracy of SPECT/CT-based lobar quantification using open-source and in-house developed segmentation models. We curated a dataset of 200 diverse CT and SPECT scans with expert lobar annotations, graded by segmentation difficulty (Easy, Moderate, Hard). We trained a local AI model on these data and included a novel anatomy-aware softmax modification at inference to improve robustness in missing-lobe cases. Lobar function error was computed as the difference between model- and ground truth-derived function after automated SPECT/CT coregistration. Results were compared across the local model, its anatomy-aware variant, three open-source models, and a commercial tool. Across 40 test cases (10 Easy, 10 Moderate, 20 Hard), no systematic bias was found in lobar function error across models. However, error standard-deviation increased significantly with case complexity for open-source models (0.57 to 5.79%, p < 0.005), particularly in missing-lobe cases (1.61 vs. 7.57%, p < 0.001). In contrast, the local model and commercial tool maintained high accuracy across difficulty (0.51 to 1.4%, p > 0.24). The anatomy-aware softmax modification eliminated false-positive lobe predictions, further improving quantification accuracy. AI-based tools can streamline routine lobar quantification with minimal but necessary physician oversight. SPECT/CT lobar quantification benefited from improved robustness through training with representative data and anatomy-awareness. While publicly available models struggled with complex cases, they produced clinically acceptable results in typical anatomy and mild diseases.