Deep learning segmentation of nasolacrimal duct on CT-DCG achieves Dice 0.79 and 84.7% classification accuracy

Automated CT-DCG analysis segmented the bony nasolacrimal duct (Dice 0.79) and classified lacrimal sac size with 84.7% agreement vs experts (κ=0.76) in a retrospective study of 151 PANDO patients.

• Retrospective single-center study of 151 patients with unilateral primary acquired nasolacrimal duct obstruction (PANDO) who underwent CT dacryocystography (CT-DCG).
• The Attention U-Net pipeline enabled automatic slice-by-slice quantification of cross-sectional area and diameters, pinpointing the narrowest plane and obstruction site.
• Limitation: Single-center retrospective design without external validation; prospective studies needed to confirm clinical utility.

RadPigeon summaries are original and for information only. They are not clinical advice.