# A Fully Automated Deep Learning Pipeline for Anatomical Landmark Localization on Three-Dimensional Pelvic Surface Scans

## TL;DR

This paper introduces an automated deep learning system for identifying pelvic landmarks in 3D scans, improving accuracy and consistency compared to manual methods.

## Contribution

A novel modular deep learning pipeline for automated pelvic landmark localization with improved repeatability and real-time performance.

## Key findings

- The system achieved a median error of 11.25 mm in landmark localization.
- Automated measurements showed higher intraclass correlation coefficients than manual marking.
- The pipeline processes scans in about three seconds, enabling near real-time clinical use.

## Abstract

Accurate identification of anatomical landmarks on three-dimensional (3D) pelvic surface scans is essential for musculoskeletal assessment, yet manual procedures remain limited by operator dependence and soft tissue variability. This study presents a fully automated deep learning pipeline for localizing anatomical landmarks on the posterior pelvic region from raw 3D point cloud data. The pipeline integrates three modules: PelvicROINet for extracting the region of interest, PelvicAlignNet for rotation correction to standardize posture, and PelvicLandmarkNet for localizing six anatomical landmarks including the bilateral posterior superior iliac spines, bilateral iliac crests, L1, and L4. The models were trained independently with task-specific annotations and combined sequentially during inference. Under a subject-level split evaluation setting, the fully integrated system achieved a median error of 11.25 mm, demonstrating consistent localization performance across unseen subjects. Compared with manual landmark marking, the automated measurements showed improved within-visit repeatability, with reduced variability and higher intraclass correlation coefficients. The entire inference process required approximately three seconds per scan, supporting near real-time clinical applicability. These results indicate that the proposed modular framework enhances numerical consistency and robustness in surface-based pelvic landmark assessment and provides a scalable foundation for AI-assisted musculoskeletal evaluation and longitudinal monitoring.