Pseudarthrosis following adult spinal deformity surgery may be predicted with preoperative MRI adipose tissue features: an artificial intelligence study on raw 3D imaging

INTRODUCTION

Pseudarthrosis (or nonunion) is defined as the failure of bone to fuse following surgical fixation and is a common complication of adult spinal deformity (ASD) surgery, with incidence rates ranging from 5%-35%^[1,2]. Pseudarthrosis is associated with recurrent pain and neurologic symptoms, can be a reason for reoperation, and can occur with or without rod fracture^[3]. Despite its prevalence and contribution to patient morbidity, the risk factors for pseudarthrosis are difficult to characterize. A preoperative risk factor is thought to be age, with multiple studies suggesting that patients over the age of 55 experience higher rates of pseudarthrosis^[4-6]. Additionally, an intraoperative risk factor is thought to be fusion to the sacrum^[7]. However, there remains debate in the literature about these risk factors, and few validated tools are available for the surgeon to preoperatively prognosticate pseudarthrosis occurrence.

Due to the difficulty in prognostication, more advanced artificial intelligence (AI) modeling techniques have been developed to augment surgical decision workflows for ASD surgery^[8-16]. Specifically, Scheer et al. developed a decision tree model from 82 variables that achieved 91% accuracy in predicting pseudarthrosis following ASD surgery^[17]. This high level of accuracy demonstrates its promise for clinical application. An underutilized extension of this framework is to utilize raw imaging data to augment predictive models. Of interest, AI models that ingest raw imaging can be directly interpreted to gain insight into nuanced patient characteristics impossible to capture in demographic variables. One imaging modality of high interest is magnetic resonance imaging (MRI) due to the detailed soft tissue signal captured. Thus, these advanced imaging models can aid preoperative decision-making, but more importantly, they can provide insight into the biological variables that may drive pseudarthrosis pathogenesis.

With the above considerations, this work aims to characterize the raw preoperative MRI features that may predict the occurrence of pseudarthrosis. In a cohort of patients undergoing ASD surgery at the major academic medical center, we sought to: (1) develop an AI model that utilizes raw preoperative MRI to predict pseudarthrosis following ASD surgery; and (2) interpret the model with class activation mapping (CAM) to understand the imaging features used to classify pseudarthrosis.

METHODS

MRI deep learning

Next, a three-dimensional convolutional neural network (3D-CNN) was developed to input raw thoracic MRIs, demographic variables, and Surgimap measured variables [Figure 1]^[18]. Only patients with MRI available were included in this study. MRI images were resliced to the three dimensions of 256 × 256 × 20 voxels, histogram equalized, and augmented using random flips, noise, bias field, blur, and affine/elastic deformations to a total of 1,080 images. Five-fold nested cross-validation with a train/validate/test split ratio of 70%/20%/10% was used to prevent overfitting and evaluate the generalizability of the model^[19]. Importantly, all splits were conducted at the patient level. A Youden index (sensitivity + specificity - 100%) was calculated for all completely withheld test partitions for each fold. The Youden index reflects the true positive, true negative, false positive, and false negative rate of the model on completely withheld validation data. The mean Youden index of all folds was tested against an equivocal null model with a Youden index value of 0 using a single population Student’s t-test^[20].

Figure 1. 3D-CNN for pseudarthrosis classification on raw thoracic MRI. 3D-CNN: Three-dimensional convolutional neural network; MRI: magnetic resonance imaging.

RESULTS

Pseudarthrosis can be predicted with raw preoperative MRI

To explore the potential of AI to non-linearly utilize the millions of data points present in raw MRI, we implemented a custom 3D-CNN. Across the five-fold nested cross-validation, the completely withheld testing data were classified by the model with a Youden index ranging from 0.30 to 0.80 (mean 0.49, 95% confidence interval ± 0.25, Figure 2). A single population t-test against a null hypothesis of a Youden index of 0.00, representing an equivocal model, was significant with a P-value of 5.50e-3. These results indicate that the 3D-CNN model was accurate in predicting pseudarthrosis following ASD surgery with at least two years of follow-up and has the potential to generalize well to a larger population.

Figure 2. Youden index across the five-fold nested cross-validation. The values shown are only for the completely withheld testing partition for each fold. ^**P < 0.01.

Superficial adipose tissue appears to be most important for classification

Of greatest interest to this work were the MRI features used by the 3D-CNN model to gain insight into the pathogenesis of pseudarthrosis. Upon model interrogation with GradCAM, the most important MRI features for classification of pseudarthrosis following ASD surgery appear to be posterior adipose tissue - i.e., the majority of GradCAM feature maps highlight various aspects of superficial adipose tissue posterior to the spinous processes (example subjects in Figure 3). Notably, there is no significant difference in body mass index (BMI) between the pseudarthrosis cohort (28.8 ± 7.3) and the non-pseudarthrosis cohort (29.0 ± 8.2), with a t-test P-value of 0.874. Notably, of the 48 patients who developed pseudarthrosis, 24 (50.0%) also had proximal junctional kyphosis (PJK). However, in our past work, we found that posterior musculature was most predictive of PJK^[18]. Thus, these results indicate that there is an important radiologic signature within these adipose regions that enables the 3D-CNN model to accurately classify pseudarthrosis, independent of total adipose content estimated by BMI and independent of radiographic features that predict PJK.

Figure 3. Posterior superficial adipose tissue appears to be important for the 3D-CNN model to classify pseudarthrosis. 3D-CNN: Three-dimensional convolutional neural network.

DISCUSSION

The current study demonstrated the accuracy of using a 3D-CNN on raw thoracic MRI to predict pseudarthrosis following ASD surgery. More importantly, the imaging features associated with pseudarthrosis were elucidated to be mainly posterior adipose tissue - with a predominance of the upper thoracic region. Interestingly, except for age, our cohort did not demonstrate any demographic or traditional radiographic measurement difference between those who developed pseudarthrosis and those who did not. Thus, it is noteworthy that the 3D-CNN heavily utilized adipose tissue of the imaging to develop the classification despite the pseudarthrosis cohort not being significantly more overweight (P = 0.874). This observation leads the authors to surmise that there exists a subtle MRI signature in the adipose tissue that the model used for classification. Furthermore, the imaging hotspots are not consistently at a region of the largest adipose collection; thus, it is likely that the 3D-CNN model is detecting an intra-adipose or adipose-adjacent signal. Future work could focus on using image segmentation techniques to better quantify the exact types of tissue present within GradCAM hotspots. Finally, posterior upper thoracic adipose tissue is typically distant from the region of pseudarthrosis, which in our cohort was predominately in the low lumbar region. Thus, it can be surmised that the network learned a global signature of pseudarthrosis predilection, not captured by demographic variables, as opposed to a local tissue abnormality at the future site of pseudarthrosis.

Overall, this work serves as an augmentation to existing models to preoperatively predict pseudarthrosis^[17]. Previously identified risk factors to predict pseudarthrosis include alcoholism, smoking, fusion location, vertebral bone quality score, diabetes, sarcopenia, advanced age, and potentially graft material^[1,23-26]. Not all of these factors were readily available in our dataset. However, it is notable that our demographic analysis of available metrics did not reveal any significant differences between pseudarthrosis and non-pseudarthrosis cohorts except for age. Thus, by incorporating raw MRI, one can leverage high-dimensional data that are often readily available in clinical databases with minimal manual extraction required. With the incorporation of readily available imaging, this technique does not rely on meticulous database mining and is free from the problems of patient bias when reporting metrics such as alcohol use and smoking status. Furthermore, this work provides potential insight into the biological underpinnings of pseudarthrosis development and could serve as background evidence for future studies exploring the role of global adipose tissue characteristics in those who develop pseudarthrosis.

It is well-documented that posterior musculature and sarcopenia appear to be strongly correlated with mechanical complications following ASD surgery^[27]. Thus, it is notable that the current work, and our past work using similar techniques to predict PJK using MRI have both revealed imaging risk factors of soft tissue - adipose tissue and posterior musculature, respectively. This contributes to the growing body of literature focusing on soft tissue characteristics as driving factors for mechanical complications, as opposed to bony anatomy. The authors do not discount the importance of bone integrity for consideration of ASD surgery, but rather aim to outline the additional importance of soft tissue health when considering a large deformity operation.

Limitations to this technique include the ever-present potential for overfitting, the computing resources required, and the technical expertise required to run the analysis algorithms. An important next step would be to test these methods on an external cohort. Another consideration with this methodology is that the thoracic MRI did not capture the top of the implanted construct in a few subjects. This can be seen as both a potential weakness and potential strength of this study because the results were robust despite this consideration - this indicates that there is possibly a global imaging feature that the 3D-CNN detects to aid accurate classification. Finally, the proper de-identification of raw data is paramount to model creation to ensure patient privacy when deploying trained models.

Overall, the use of machine learning in medical imaging has garnered attention but has still been limited in scope compared to tabular data machine learning and large language models. We aimed to demonstrate the potential of a simple classification scheme on available 3D MRIs to predict the development of pseudarthrosis following ASD surgery. Beyond the cross-validated accuracy of the model, our approach has the benefit of providing a level of interpretation by outlining imaging features used by the model to make classification decisions. Overall, this work demonstrates the capabilities of raw imaging AI in spine surgery and can serve as the basis for a deeper biological inquiry into the pathogenesis of pseudarthrosis.