Pedvision: A Manual-Annotation-Free and Age Scalable Segmentation Pipeline for Bone Analysis in Hand X-Ray Images

26 Pages Posted: 18 Dec 2024

See all articles by Morteza Homayounfar

Morteza Homayounfar

Delft University of Technology (TU Delft); Erasmus University Rotterdam (EUR) - Erasmus Medical Center (MC)

Sita M.A. Bierma-Zeinstra

Erasmus University Rotterdam (EUR) - Department of General Practice

Amir A. Zadpoor

Delft University of Technology - Department of Biomechanical Engineering

Nazli Tümer

Delft University of Technology

Abstract

Medical image analysis often involves time-consuming annotation processes. Pediatric image analysis adds further complexity due to the scarcity of data, noise, and growth-related anatomical variations, especially in bone analysis, where bone structures take longer to evolve compared to other organs. This study aims to develop a segmentation model that scales across different age groups, reduces annotation effort, and ensures high accuracy, particularly in low-quality images.To address these challenges, we propose a segmentation pipeline that first uses a Region of Interest (ROI) network to identify relevant regions, followed by a foundation model that translates each region into meaningful instances. These instances are then mapped to segmentation classes through an instance classifier network. To initiate rounds of the training of ROI and instance classifier models, we developed a fast, semi-automated annotation framework that leverages foundation models to annotate a subset of images using an object-level approach. In subsequent rounds, a human discriminator selects promising predictions of the last round fed by unseen data, progressively enriching the model’s training dataset for further fine-tuning of networks. The networks are expanded from low-parameter to high-parameter models across rounds, incorporating a curriculum learning approach to capture increasingly complex features.We evaluated our pipeline on hand X-ray images from children aged 4 to 228 months, segmenting 19 bones into five classes. The pipeline results compared to DeepLabV3+ models with ResNet34 and ResNet101 backbones, achieving up to a 6.19% improvement in Dice scores for subjects over 14 years and 17.01% for those under 7, particularly in low-quality images. An ablation study is performed to explore pipeline robustness against parameter and configuration variations. The project is open source at github.com/mohofar/PedVision.

Note:
Funding Information: Funding details are included in the manuscript (European Research Council, grant nr. ERC-adv. 101054778)

Conflict of Interests: The authors have no competing interests to declare.

Keywords: Pediatric image analysis, Visual foundation model, Hand bone segmentation, X-rays, Manual-annotation-free, Human-in-the-loop training.

Suggested Citation

Homayounfar, Morteza and Bierma-Zeinstra, Sita M.A. and Zadpoor, Amir A. and Tümer, Nazli, Pedvision: A Manual-Annotation-Free and Age Scalable Segmentation Pipeline for Bone Analysis in Hand X-Ray Images. Available at SSRN: https://ssrn.com/abstract=5050535 or http://dx.doi.org/10.2139/ssrn.5050535

Morteza Homayounfar (Contact Author)

Delft University of Technology (TU Delft) ( email )

Jaffalaan 5
Delft, 2628BX
Netherlands

Erasmus University Rotterdam (EUR) - Erasmus Medical Center (MC) ( email )

Doctor Molewaterplein 40
Rotterdam, South Holland 3015 GD
Netherlands

Sita M.A. Bierma-Zeinstra

Erasmus University Rotterdam (EUR) - Department of General Practice ( email )

Amir A. Zadpoor

Delft University of Technology - Department of Biomechanical Engineering ( email )

Delft
Netherlands

Nazli Tümer

Delft University of Technology ( email )

Stevinweg 1
Stevinweg 1
Delft, 2628 CN
Netherlands

Do you have a job opening that you would like to promote on SSRN?

Paper statistics

Downloads
24
Abstract Views
85
PlumX Metrics