Cascade learning in multi-task encoder-decoder networks for concurrent bone segmentation and glenohumeral joint assessment in shoulder CT scans

TL;DR

This work introduces an innovative deep-learning framework processing shoulder CT scans that aims to streamline the preoperative planning pipeline delivering high-quality bone surfaces and supporting surgeons in selecting the most suitable surgical approach according to the unique patient joint conditions.

Abstract

Osteoarthritis is a degenerative condition affecting bones and cartilage, often leading to osteophyte formation, bone density loss, and joint space narrowing. Treatment options to restore normal joint function vary depending on the severity of the condition. This work introduces an innovative deep-learning framework processing shoulder CT scans. It features the semantic segmentation of the proximal humerus and scapula, the 3D reconstruction of bone surfaces, the identification of the glenohumeral (GH) joint region, and the staging of three common osteoarthritic-related pathologies: osteophyte formation (OS), GH space reduction (JS), and humeroscapular alignment (HSA). The pipeline comprises two cascaded CNN architectures: 3D CEL-UNet for segmentation and 3D Arthro-Net for threefold classification. A retrospective dataset of 571 CT scans featuring patients with various degrees of GH osteoarthritic-related pathologies was used to train, validate, and test the pipeline. Root mean squared error and Hausdorff distance median values for 3D reconstruction were 0.22mm and 1.48mm for the humerus and 0.24mm and 1.48mm for the scapula, outperforming state-of-the-art architectures and making it potentially suitable for a PSI-based shoulder arthroplasty preoperative plan context. The classification accuracy for OS, JS, and HSA consistently reached around 90% across all three categories. The computational time for the inference pipeline was less than 15s, showcasing the framework's efficiency and compatibility with orthopedic radiology practice. The outcomes represent a promising advancement toward the medical translation of artificial intelligence tools. This progress aims to streamline the preoperative planning pipeline delivering high-quality bone surfaces and supporting surgeons in selecting the most suitable surgical approach according to the unique patient joint conditions.

Figures (11)

• Figure 1: The proposed processing pipeline involving the volume segmentation to extract the humerus and the scapula (CEL-UNet model), the detection of the humeral head region in the original CT volume exploiting the two reconstructed surfaces, and finally the multiclass staging of the GH joint (Arthro-Net model), in terms of osteophyte staging (3 classes of increasing osteophyte size), GH joint space (3 classes of increasing severity) and humeral scapular alignment (2 classes).
• Figure 2: Example of bone volumes provided in the dataset A) original humerus morphology with osteophytes pointed out by the red arrow B) osteophyte-cleared humerus C) scapula.
• Figure 3: Segmentation (DSeg) and classification (DClass) training set generation pipeline. After CT normalization, DSeg is produced by cropping the volume to focus on the shoulder joint and patching the extracted sub-volumes, while for DClass a GH-centered crop was performed, and data augmentation was obtained by flipping the sub-volumes on the sagittal plane.
• Figure 4: Different pathological conditions. A) Extreme eccentricity of the humeral head, B) Mild GH osteoarthritis, C) Severe GH osteoarthritis with large osteophyte presence.
• Figure 5: CEL-UNet architecture. The encoding path, on the left, extracts segmentation features at different resolutions, ending with a bottleneck. The decoder, on the right, is split into two parallel branches, devoted to region segmentation (RA) and edge detection (CA). The output of each CA decoding block is combined with the corresponding RA layer through vertical skip connections (red lines) to enhance semantic segmentation. Pyramidal edge extraction (PEE) modules (yellow boxes) are added after decoding blocks. The segmentation mask output is then used to reconstruct the bone surfaces using traditional matching cubes algorithm.