Deep learning-based segmentation and quantification of pulmonary masses on apparent diffusion coefficient maps: a multicentre reproducibility study

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: Internal DSC: 0.843; External DSC: 0.701 (centres 2 and 3 combined or separately not specified). Mean ADC Bland-Altman mean difference: 0.00 × 10⁻³ mm²/s (model vs. senior radiologist). RDC for mean ADC: 12.98% (model-senior) vs. 22.87% (senior-junior radiologist). Test-retest CCC: 0.931 (ADCSegNet) vs. 0.908 (radiologists)

• Primary Outcome: Segmentation accuracy of ADCSegNet on pulmonary ADC maps, assessed by Dice similarity coefficient (DSC), and reproducibility of automated ADC quantification compared with manual radiologist measurements

• Nnt Or Sensitivity: DSC (segmentation accuracy): 0.843 internal, 0.701 external. ICC(2,1) >0.75 for most ADC metrics. Mean ADC RDC%: 12.98% (model-senior). Test-retest CCC: 0.931 (model) vs. 0.908 (radiologists)

• Confidence Interval: Bland-Altman 95% limits of agreement for mean ADC: -0.13 to 0.13 × 10⁻³ mm²/s (model vs. senior radiologist). Confidence intervals for DSC and CCC not reported in abstract

Clinical Application

Feasibility in routine clinical practice is currently limited. DWI-MRI for pulmonary mass characterisation is not yet standard of care in most centres. The model requires ADC map inputs generated from specific MRI protocols, and integration into clinical PACS or reporting workflows is not described. The external DSC of 0.701 at centre 3 (n=12) suggests performance degradation with protocol variation, which would require site-specific validation before clinical deployment. Pulmonary MRI with DWI is not currently a standard diagnostic pathway in Australian practice for lung mass evaluation; CT and PET-CT remain first-line modalities per RACGP and thoracic oncology guidelines. The TGA has not approved ADCSegNet or equivalent tools for clinical use in Australia. PBS does not fund pulmonary DWI-MRI as a routine investigation. Australian adoption would require prospective validation in Australian populations, TGA regulatory clearance as a software as a medical device (SaMD), and demonstration of clinical utility over existing CT-based workflows. The RANZCR may have interest in this technology for future guideline development, but current evidence is insufficient to support practice change. Adult patients with pulmonary masses (likely including both benign and malignant lesions) undergoing diffusion-weighted MRI with ADC map generation at tertiary radiology centres with research-grade MRI capability

Abstract

OBJECTIVES: 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. METHODS: 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). RESULTS: 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) × 10⁻3 mm2/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). CONCLUSIONS: 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.

References

1. 1.Ding, L., Chen, T., Gao, Q., Liu, P., Yang, M., Xu, H., Liu, J., Zhou, J., Wan, Q., & Li, S. (2026). Deep learning-based segmentation and quantification of pulmonary masses on apparent diffusion coefficient maps: a multicentre reproducibility study. European Journal of Radiology. https://doi.org/10.1016/j.ejrad.2026.112881