Exploratory deep learning applications for orthopaedics clinical practices in spinal disorders

Abstract

With the development of computer science and technology, deep learning has been widely applied in medical clinical scenarios. There exist challenges in orthopaedic practice, especially for Idiopathic Scoliosis (IS) and Cervical Spondylotic Myelopathy (CSM). Currently, there are several well-established procedures in orthopaedic practice. However, some procedures or solutions to clinical problems are often constrained by specialized knowledge, which can make them time-consuming and laborious. Most existing analysis methods focus on features from a single aspect, thereby neglecting interactions among different features. These limitations can reduce efficiency or potentially hinder problem-solving capabilities. While deep learning methods hold promise for addressing these issues, appropriate deep learning approaches can extract deeper and more abstract features from a comprehensive perspective, thereby addressing complex clinical challenges associated with orthopaedic practice, such as diagnosis, segmentation, and prognosis, in a more efficient manner. Therefore, this project aims to explore the feasibility and efficacy of deep learning approaches in addressing clinical challenges in orthopaedic practice.

Four projects were undertaken covering various clinical settings and diverse data formats of different spinal disorders including: 1. Somatosensory evoked potentials (SEP) signal prediction for intraoperative monitoring. Traditional methods based on static thresholds are prone to generating false alarms due to non-injury factors that affect SEP signal amplitudes. By leveraging the strengths of long short-term memory and convolutional neural networks, a prediction model for SEP amplitude during scoliosis surgery has been developed. This model can accurately quantify changes in intraoperative SEP amplitude under normal conditions using physiological signals. 2. Segmentation of regions of interest within the spinal cord for CSM patients. Traditional segmentation methods focus on manual segmentation of gray matter and white matter or the whole spinal cord. A deep learning model with U-Net architecture combined with a proposed loss function based on heatmap distance achieves segmentation of bilateral ventral, lateral, and dorsal columns, as well as gray matter within the spinal cord in diffusion tensor imaging (DTI) images. 3. Prognosis of CSM based on DTI metrics and clinical data. Current studies primarily focus on analyzing individual features of spinal DTI images, but they fail to fully explore the interactions among features. The proposed method, through end-to-end training, comprehensively analyzes the complex interactions between DTI metrics features and clinical features achieving more accurate prediction of surgical outcomes. 4. Skeletal maturity assessment based on Proximal Femur Maturity Index (PFMI). Traditional criteria for evaluating skeletal maturity exhibit low sensitivity or pose a risk of radiation exposure. PFMI is a new criterion aided by biplanar stereoradiography, which mitigates the risk of radiation exposure. However, PFMI still relies on professional knowledge, which limits its widespread adoption. Deep learning models have demonstrated superior performance in bone maturity evaluation and are suitable for deployment in real-time applications and resource-limited settings due to their compressed model parameters and rapid inference times.

Deep learning methods have demonstrated satisfactory performance in these clinical applications. Existing evidence suggests that deep learning approaches are both feasible and effective for addressing clinical challenges in orthopaedic practice.