Deep Learning-Based Automated Segmentation and Multi-Parametric Quantitative Assessment of the Lacrimal Drainage System on CT-DCG.
Authors
Affiliations (3)
Affiliations (3)
- Eye Center, The Second Affiliated Hospital, School of Medicine, Zhejiang University, Zhejiang Provincial Key Laboratory of Ophthalmology, Zhejiang Provincial Clinical Research Center for Eye Diseases, Zhejiang Provincial Engineering Institute on Eye Diseases, Hangzhou, Zhejiang, People's Republic of China.
- Department of Plastic Surgery, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, People's Republic of China.
- Ningbo Key Laboratory of Medical Research on Blinding Eye Diseases, Ningbo Eye Institute, Ningbo Eye Hospital, Wenzhou Medical University, Ningbo, People's Republic of China.
Abstract
Primary acquired nasolacrimal duct obstruction (PANDO) is a common cause of epiphora. Because surgical outcomes are highly dependent on the obstruction site and anatomic variations, this study presents a deep learning-based framework for the automated morphological and quantitative analysis of the lacrimal system via CT dacryocystography (CT-DCG). A total of 151 patients with unilateral PANDO who underwent CT-DCG were included. An automated pipeline utilizing a deep learning network (Attention U-Net) was developed to segment and reconstruct the lacrimal sac and bony nasolacrimal duct (BNLD). The system performed automated quantification of the long axis, short axis, and cross-sectional area for each slice from the superior to the inferior opening of the BNLD, as well as for the lacrimal sac. Segmentation accuracy was evaluated using the Dice coefficient, and the clinical agreement of lacrimal sac size classification was assessed using Cohen's kappa coefficient compared to expert manual assessment. The deep learning model achieved robust performance in segmenting the lacrimal drainage system, with an average Dice coefficient of 0.79 for the BNLD. The system successfully enabled automatic, slice-by-slice morphological quantification, providing continuous measurement curves for the cross-sectional area and diameters throughout the entire duct. It accurately identified and localized the narrowest cross-sectional plane and the obstruction site within the BNLD. In the classification of lacrimal sac size, the automated morphological analysis demonstrated high consistency with expert assessment, achieving an overall agreement rate of 84.7% and a Cohen's kappa coefficient of 0.76. The proposed deep learning framework demonstrates the technical feasibility of automated segmentation and morphological quantification of the lacrimal sac and BNLD on CT-DCG. Although further prospective validation is required, the objective anatomic metrics provided by this system hold promising potential to assist in preoperative assessment and surgical planning for PANDO treatment. This proof-of-concept study demonstrates that automating the segmentation and quantification of the lacrimal drainage system on CT-DCG can provide an efficient, objective alternative to manual assessment. Whereas its definitive impact on surgical outcomes remains to be established, the system yields standardized morphological metrics that have the potential to support individualized preoperative evaluation and clinical decision making for nasolacrimal duct obstruction.