Clinicians must promptly and accurately distinguish acute VCFs from normal vertebrae, and chronic VCFs to facilitate early evaluation of the surgical strategy and determine the etiology of the acute VCFs. Although modern MRI techniques have successfully addressed the challenges of locating and identifying acute VCFs, patients often endure lengthy delays in undergoing MRI, resulting in pain and discomfort during the waiting period. The results of this study describe the development of a novel multi-scene DL algorithm specifically designed to segment acute VCFs within spine radiographs. We further confirmed the ability of this innovation through meticulous internal and external validation using a comprehensive multi-institutional dataset. As a multi-scene segmentation model, we rigorously evaluated the PFNet in both preoperative and intraoperative radiographs, demonstrating its performance and effectiveness in distinct clinical scenarios. Our comparative analysis against competitive models unequivocally demonstrated the consistent superiority of the PFNet, with unparalleled performance across all datasets and metrics. These findings confirmed that the PFNet is a leading model that excelled in accurately segmenting acute VCFs across diverse imaging scenarios.

The high degree of similarity between acute VCFs and normal vertebral bodies in radiographs [14] can lead to false positive and false negative predictions in the initial segmentation results. To address these erroneous predictions, we designed an SDM, which takes current-level features, higher-level features, and the predicted results as inputs and outputs refined features and more accurate predictions. This module aims to refine the segmentation results by leveraging contextual information from multiple levels and incorporating feedback from higher-level features, ultimately improving the segmentation accuracy. Additionally, the AGM consists of channel attention blocks and spatial attention blocks. Both blocks employ non-local operations to capture long-range dependencies between channels and spatial positions, thereby enhancing the semantic representation of the deepest layer features from a global perspective.

Unlike other DL models that have limited image availability for training and external validation [5, 14], our PFNet was developed using an extensive dataset obtained from two hospitals. Furthermore, we validated the robustness and reliability of the PFNet using three distinct and independent datasets from three separate hospitals. This comprehensive dataset and thorough validation support the credibility and generalizability of the PFNet, distinguishing it from models that are limited by data availability and validation. Additionally, the limited external generalizability of DL models significantly hinders their translation into clinical practice [7]. The radiographs used in this study were generated by various instruments, ensuring the reproducibility of the PFNet model’s excellent performance [23, 24]. These findings highlight the significant improvements made by the PFNet, demonstrating its superior performance not only within the validation dataset but also across multiple external test datasets, indicating its exceptional generalization and consistently superior performance.

To the best of our knowledge, the PFNet model is the first DL model applied to segmenting acute VCFs from intraoperative radiographs. Surgeons rely solely on bone markers in radiographs to identify acute VCFs intraoperatively. Consequently, the condition may be confused with diseased vertebral bodies, resulting in serious medical consequences. Our analysis of the C-arm fluoroscopy dataset demonstrated the efficacy of the PFNet as a multi-scene segmentation DL model in analyzing intraoperative radiographs. However, the sensitivity of the PFNet model did not meet expectations, unlike its success in other performance metrics. This performance discrepancy between preoperative and intraoperative scenarios can be attributed to several factors. Firstly, the limited availability of intraoperative training data and the resulting domain shift could significantly impact the model’s performance. Additionally, insufficient training data specific to intraoperative tasks likely contributed to its suboptimal performance. Moreover, various surgical factors and instruments, such as tissue manipulation and the presence of metal implants, can complicate image interpretation. These factors could have influenced the model’s performance in interpreting intraoperative radiographs, thus impacting its sensitivity. In conclusion, the limited availability of intraoperative radiographs for training, coupled with insufficient data for the unique challenges posed by intraoperative scenarios and potential confounding surgical factors, collectively contribute to the suboptimal sensitivity observed in the performance of the PFNet within this context.

Accurately segmenting anatomical structures within medical images presents a crucial challenge for clinical diagnosis and analysis [25, 26]. In clinical settings, clinicians and radiologists dedicate significant time and effort to segmenting and delineating regions of interest within medical images. Therefore, the accurate and automated segmentation of medical images can alleviate the laborious and time-consuming burden on radiologists [27, 28].

The operational workflow of our PFNet model involves two main aspects: identifying acute VCFs and segmenting fractured vertebral bodies. Notably, previous studies have also developed DL models for segmenting anatomical structures in radiological images [29,30,31,32]. However, these models largely focused on normal or healthy anatomical structures, potentially falling short in effectively addressing clinical needs for disease diagnosis. Furthermore, previous studies have misunderstood the identification of acute VCFs as a simple classification task. This has led to models that lack interpretability and hinder their clinical usefulness. Previously, diagnosing and detecting acute VCFs in spinal radiographs required segmenting each vertebral body from the entire spine [5]. However, conventional models that combined classification and segmentation modules were often cumbersome and computationally complex. This could result in potential feature loss or overfitting. In contrast, our PFNet model excels in directly segmenting acute VCFs, thereby improving the clinical workflow and enhancing the efficiency of segmentation and diagnosis.

The PFNet model has some limitations. Firstly, although we have implemented the model in intraoperative radiographs, obtaining a substantial number of these images for training purposes remains challenging. Therefore, a larger and more diverse training set of intraoperative radiographs is needed to ensure the robustness and stability of the model. Secondly, owing to the retrospective study design, selection bias and inherent differences were inevitable. Therefore, prospective trials are needed and validation in prospective datasets is necessary to support the conclusions of this study. Additionally, we speculate that the bone density, fracture severity, degree of vertebral compression, and image quality may affect our model’s performance. As shown above, the PFNet model did not achieve high sensitivity in the C-arm fluoroscopy dataset. These issues may be attributed to factors such as insufficient data, low-quality images, and the complexity of specific patterns in C-arm fluoroscopy, among others.

To address these limitations, future work should focus on optimizing the algorithm through various means. Potential optimization tools include using image enhancement techniques (such as Retinex [33]) to preprocess the dataset to further improve segmentation accuracy. Additionally, incorporating mathematical models to constrain and guide the prediction results, such as the alternating direction method of multipliers algorithm, would provide a stronger mathematical basis for the model during prediction. Furthermore, continued testing and validation across diverse datasets will be essential to refine the algorithm’s accuracy and ensure its robustness in different practical applications. By addressing these limitations and exploring optimization avenues, the algorithm’s potential impact on decision-making processes can be fully realized, leading to more reliable and efficient outcomes in everyday practice.