Deep learning-based segmentation and quantification of pulmonary masses on apparent diffusion coefficient maps: a multicentre reproducibility study.
Authors
Affiliations (5)
Affiliations (5)
- School of Biomedical Engineering, First Affiliated Hospital of Guangzhou Medical University, Guangzhou 511436, China; School of Communication and Electronic Engineering, Jishou University, Jishou 416000, China.
- Department of Radiology, National Center for Respiratory Medicine, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou 510120, China.
- School of Biomedical Engineering, First Affiliated Hospital of Guangzhou Medical University, Guangzhou 511436, China.
- Department of Radiology, National Center for Respiratory Medicine, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou 510120, China. Electronic address: [email protected].
- School of Biomedical Engineering, First Affiliated Hospital of Guangzhou Medical University, Guangzhou 511436, China. Electronic address: [email protected].
Abstract
To develop and validate a deep learning model for automatic segmentation of pulmonary masses on apparent diffusion coefficient (ADC) maps and to assess repeatability of automated ADC quantification. We proposed ADCSegNet, a deep learning model tailored to pulmonary ADC maps, trained and tested on ADC maps from centre 1 (303 MRI examinations) and externally evaluated on datasets from centre 2 (70 examinations) and centre 3 (12 examinations) for generalisability. Segmentation accuracy was evaluated using the Dice similarity coefficient (DSC). Agreement between model-derived and manual ADC measurements (senior and junior radiologists) was examined using Bland-Altman analysis, the intraclass correlation coefficient [ICC(2,1)], and the reproducibility coefficient (RDC) following QIBA profiles. Repeat-scan stability was assessed on an independent prospective test-retest dataset (22 × 2 examinations) using Lin's concordance correlation coefficient (CCC). ADCSegNet achieved DSCs of 0.843 internally and 0.701 externally, outperforming nnU-Net and other contemporary backbones. Agreement was highest between the model and the senior radiologist; for mean ADC, the mean difference was 0.00 (95% limits of agreement, -0.13 to 0.13) × 10⁻<sup>3</sup> mm<sup>2</sup>/s, exceeding model-junior and inter-radiologist agreement. ICC(2,1) exceeded 0.75 for most ADC metrics, and RDC confirmed high reproducibility for key metrics (mean ADC RDC%: 12.98% for model-senior vs. 22.87% for senior-junior radiologist). Test-retest analyses showed higher repeatability of ADCSegNet than radiologists for core metrics such as mean ADC (CCC 0.931 vs. 0.908). ADCSegNet enables accurate segmentation of pulmonary masses on ADC maps and provides reproducible automated ADC quantification, outperforming state-of-the-art backbone models while showing strong agreement with expert readers and stable performance on repeat scans.