Transforming Bone Tunnel Evaluation in Anterior Cruciate Ligament Reconstruction: Introducing a Novel Deep Learning System and the TB-Seg Dataset

Abstract

Evaluating bone tunnels is crucial for assessing functional recovery after anterior cruciate ligament reconstruction. Conventional methods are imprecise, time-consuming, and labor-intensive. This study introduces a novel deep learning-based system for accurate bone tunnel segmentation and assessment. The system has two primary stages. Firstly, the ResNet50-Unet network is employed to capture the bone tunnel area in each slice. Subsequently, in the bone texture analysis, the open-source software 3D Slicer is leveraged to execute three-dimensional reconstruction based on the segmented outcomes from the previous stage. The ResNet50-Unet network was trained and validated using a newly developed dataset named tunnel bone segmentation (TB-Seg). The outcomes reveal commendable performance metrics, with mean intersection over union (mIoU), mean average precision (mAP), precision, and recall on the validation set reaching 76%, 85%, 88%, and 85%, respectively. To assess the robustness of our innovative bone texture system, we conducted tests on a cohort of 24 patients, successfully extracting bone volume/total volume, trabecular thickness, trabecular separation, trabecular number, and volumetric information. The system excels with substantial significance in facilitating the subsequent analysis of the intricate interplay between bone tunnel characteristics and the postoperative recovery trajectory after anterior cruciate ligament reconstruction. Furthermore, in our five randomly selected cases, clinicians utilizing our system completed the entire analytical workflow in a mere 357–429 s, representing a substantial improvement compared to the conventional duration exceeding one hour.