HR-pQCT-based Joint Analysis

Automatic 3D Joint Erosion and Bone Proliferation Detection for the Diagnosis and Monitoring of Rheumatoid Arthritis Using Hand HR-pQCT Images

Code (v1) available at: HR-pQCT-joint-source.zip

Algorithm Overview

The proposed framework consists of three main parts:

Part I: Cortical surface reconstruction and atlas construction for population-level healthy joint modeling
Part II: Elastic shape model to estimate healthy cortical surface from diseased joints
Part III: Quantitative filtration and visualization of bone proliferation regions

The complete workflow is illustrated in Figure 1 (cortical reconstruction → atlas-based shape analysis → proliferation detection and visualization).

Workflow of bone proliferation detection algorithm — Figure 1. The workflow of bone proliferation detection. The first step (in the red dashed pane) constructs atlas for population-level analysis. The second step (in the blue dashed pane) detects joint proliferation using shape analysis methods. The third step (in the green dashed pane) filtrates proliferation regions (best viewed in color).

Part I: Atlas Construction

Based on previous cortical surface reconstruction work, enhanced with 3D nnU-Net segmentation and level-set refinement using GPU AutoDiff acceleration
nnU-Net trained on physician-annotated joint volume masks from CUHK
Offline atlas (canonical healthy joint surface) constructed using Karcher mean estimation under elastic shape distance
Displacement fields parameterized by SIREN neural network

Part II: Shape Analysis-based Joint Proliferation Detection

Core technique: Elastic deformation model with neural displacement field parameterization (SIREN network).

Elastic distance defined with multiple energy terms (default parameters: 200, 200, 100, 20)
Optimized using PyTorch + Adam + ReduceLROnPlateau scheduler (initial lr = 1e-4)
Atlas estimation formulated as Karcher mean problem with rigid alignment preprocessing
Algorithm 1 describes the iterative random sampling and displacement field update procedure

Part III: Quantitative Proliferation Detection

Rigid alignment of diseased surface to atlas → compute displacement field → deform atlas
Calculate point-to-surface (P2S) distance (positive = outward protrusion)
Threshold P2S distance to identify coherent proliferation regions
Cluster using DBSCAN
Reconstruct watertight surfaces with 3D alpha shape algorithm
Filter out tiny-volume meshes to retain significant proliferations

Validation on Simulated Data

Healthy joint surfaces polished as ground truth → synthetic protrusions manually created in Blender → certified by physicians.

20 paired simulated samples generated (covering common Type I & II proliferations simulation_diseased_bone.7z simulation_GT.7z).