Establishment and Reliability of an Automatic Measurement Method of Pectus Excavatum Indices Using a Deep Learning Model

TL;DR

This study developed a deep learning model to automatically measure pectus excavatum severity indices, showing high reliability compared to manual methods.

Contribution

The novel contribution is a U-Net-based automated system for measuring pectus excavatum indices with high consistency and accuracy.

Findings

Manual measurements had a 15.9% error rate, reduced to 4.1% after consensus correction.

The U-Net model showed stable error rates (8.7% vs 8.5%) with strong agreement (ICC 0.83-0.92) to corrected manual measurements.

Bland-Altman analysis confirmed minimal bias between automated and manual measurements.

Abstract

Objective: This study aimed to evaluate the consistency and accuracy of pectus excavatum (PE) indices assessment by comparing U-Net-based automated segmentation with manual measurements, aiming to reduce interobserver variability and standardize clinical workflow in PE severity evaluation.

Methods: An automatic measurement model was developed using U-Net architecture, trained on 550 chest computed tomography (CT) scans from 94 patients and validated on 164 independent scans. The model calculated three key indices (Haller, correction, and asymmetry), compared against measurements by four observers.

Results: Manual measurements showed an initial error rate of 15.9% (first three observers), reduced to 4.1% after consensus correction (p<0.01). The U-Net model exhibited stable error rates (8.7% vs 8.5% pre-/post-correction, p=0.91). Strong agreement was observed between automated and corrected manual measurements: Haller index (intra-class correlation (ICC)=0.83), correction index (ICC=0.86), asymmetry index (ICC=0.92) (all p<0.01). Bland-Altman analysis confirmed minimal bias.

Conclusion: U-Net-based automation provides reliable measurement of PE severity indices, demonstrating the potential to reduce observer-dependent variability and enhance clinical workflow efficiency. Multi-center validation is warranted to support broader radiologic applications.

Figures