Garden classification of femoral neck fracture using deep-learning algorithm

Comprehensive Summary

In this study, Yang et al. used a deep-learning approach to create a novel method of assessing X-ray images of femoral neck fractures (FNFs) to avoid the traditional CT-based approach of Garden classification score prediction. Data was collected from 1588 patients, comprising antero-posterior (AP) and lateral (LA) view X-rays, and pre-operative 3D CT-scans, all taken within 2 days of hospital admission for FNF. A 2-step process using independent convolutional neural networks (CNNs) was employed, in which the first step identified and processed images of hip joints across all 3 image views, and the second step processed all images. Reference standards used CT scans, with an orthopedic professor confirming 100 patients. The models developed were able to effectively identify the hip joint with around 98.5 - 100% precision and recall. Identification of non-displaced FNFs (Garden type I&II) had an accuracy of 87.2 - 89.6%, and displaced (Garden type III&IV) had maximum precision and recall values of 94.8% and 91.9%, respectively. Among the models used, the highest Area-under-the-curve (AUC) value of 0.93 was MobileNetV3. Overall, the data show comparable performance to that of CT imaging, and external validations show a 88.4% accuracy for the ensemble model.

Outcomes and Implications

Unlike previous studies with disparate results for femoral neck injury classification, this study presents a reliable and accurate deep-learning model for the assessment of FNFs with the Garden classification system. FNFs have a high likelihood of osteonecrosis due to the chance of vascular injury, but decision-making on what surgery procedure to take requires a less-than-objective method of classification due to the low inter-observer reliability of the Garden classification between clinicians. In comparison to previous studies with low reliability or current methodology, this study presents a lightweight, computationally efficient model with a higher rate of accuracy that can also use less data by relying on X-ray scans. Given more time to improve accuracy and further generalizability by using multi-center datasets, this approach of deep learning can serve as a valuable diagnostic tool.