Space channel enhanced diffusion model-based scoliosis postoperative prediction method

Abstract

The invention discloses a scoliosis postoperative prediction method based on a space channel enhanced diffusion model, which comprises the following steps of: collecting preoperative and postoperative X-ray images into a database, labeling and preprocessing the preoperative and postoperative X-ray images, extracting a region of interest, carrying out target detection, and then processing all data sets; and finally, generating an output matched with the preoperative X-ray image by using a diffusion model based on spatial channel enhancement. According to the method, an intelligent mapping model from preoperative images to postoperative orthopedic effect prediction is constructed by combining a target detection technology and a generative artificial intelligence technology. According to the technical scheme, a quantitative prediction tool can be provided for scoliosis orthopedic surgery, clinical doctors are assisted in surgical plan planning and prognosis evaluation, and personalized precise medical decision making is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical artificial intelligence and computer-aided diagnosis, and specifically relates to a method for postoperative prediction of scoliosis based on a spatial channel enhanced diffusion model. Background Art

[0002] Adolescent idiopathic scoliosis (AIS) is a three-dimensional spinal deformity affecting individuals aged 10 to 18 years with a curvature of 10° or greater. AIS is the most common spinal disorder in adolescents, affecting 1% to 4% of the population, with a greater prevalence among adolescent girls. Its etiology remains unclear. Mild scoliosis is often asymptomatic, while progressive moderate to severe scoliosis can cause back deformity, respiratory impairment, paralysis, and even life-threatening conditions. For patients with progressive AIS who fail conservative treatment, corrective surgery is virtually the only treatment option. However, there is no consensus on the optimal fusion level. Improper fusion level selection can lead to postoperative complications, including coronal decompensation, progression of the primary or secondary curve, proximal or distal junctional kyphosis, and internal fixation failure. Severe cases may even require revision surgery.

[0003] Therefore, accurately predicting the outcome of scoliosis surgery is crucial. This approach not only helps doctors develop scientific, personalized treatment plans, but also significantly increases surgical success rates, reduces complications, and ultimately improves patients' quality of life. However, in current clinical practice, surgical planning primarily relies on the physician's personal experience and intuition. This approach struggles to fully account for individual patient differences and the complexity of the surgery, and is inherently subjective and limited.

[0004] Predicting postoperative orthopedic outcomes remains a relatively understudied area, yet this prediction is crucial for selecting the optimal surgical strategy and assessing postoperative prognosis. By integrating patients' preoperative and postoperative imaging data, deep learning techniques can provide strong support for personalized orthopedic prediction. These models not only provide more accurate decision support but also, through in-depth analysis of big data, identify potential risk factors, thereby helping physicians develop more personalized and effective treatment plans. Currently, deep learning-based prediction models have an accuracy rate exceeding 80% in predicting postoperative orthopedic success. This result not only outperforms traditional statistical prediction methods but also provides more detailed information on the potential recovery trajectory of individual patients. However, current methods lack intuitiveness and rely primarily on data fitting. We hope to be able to directly predict postoperative outcomes through images.

[0005] To address these issues, we propose a deep learning-based solution: a diffusion model based on spatial channel enhancement for postoperative scoliosis prediction. Diffusion models are a type of generative model based on probabilistic modeling. Their core concept is to gradually add noise to an image during training until it approximates pure Gaussian noise, then gradually reconstruct the original image through a learned denoising process. Unlike traditional generative adversarial networks (GANs), diffusion models do not rely on adversarial training, making them more stable and capable of generating high-quality and diverse images. Basic diffusion models directly model images in pixel space, which, while offering excellent performance, have limitations in practical applications due to their high computational overhead and slow generation speed. To address this issue, we employ a latent diffusion model. This model first compresses the image into a latent space using a pretrained autoencoder (such as VQ-VAE or VAE) and then performs diffusion modeling on this low-dimensional latent representation. This significantly improves computational efficiency and generation speed while preserving image quality and semantic consistency. Our proposed method, based on the design concept of a conditional latent diffusion model, uses preoperative images as conditional inputs to guide the model in learning the correspondence between preoperative and postoperative images during the generative process. This conditional generative mechanism effectively improves the relevance and medical usability of the generated results. Ultimately, our method can directly predict postoperative outcomes, providing a more intuitive and comprehensive reference for AIS surgical planning, reducing reliance on subjective experience and further improving the scientific nature and quality of surgical planning. Summary of the Invention

[0006] Purpose of the invention: The present invention provides a method for postoperative prediction of scoliosis based on a diffusion model enhanced by spatial channels, so as to solve the problems raised in the above-mentioned background technology. The present invention collects preoperative and postoperative X-ray images into a database, annotates and preprocesses them, extracts regions of interest, performs target detection, and then processes all data sets, and finally uses a diffusion model based on spatial channel enhancement to generate an output that matches the preoperative X-ray image. The present invention combines target detection technology with generative artificial intelligence technology to construct an intelligent mapping model from preoperative images to postoperative correction effect prediction. The technical solution of the present invention can provide a quantitative prediction tool for scoliosis correction surgery, assist clinicians in surgical plan planning and prognosis evaluation, and realize personalized precision medical decision-making.

[0007] The technical solution adopted by the present invention is a method for predicting scoliosis after surgery based on a spatial channel enhanced diffusion model, comprising the following steps: Step 1: Training Data Collection, Labeling, and Preprocessing: Preoperative and postoperative X-ray images of adolescent idiopathic scoliosis (AIS) patients were collected to create a database. The images in the database were labeled and preprocessed to extract the complete spinal-pelvic region from paired preoperative and postoperative X-ray images, covering the atlanto-occipital joint to the pubic symphysis. Step 2: Extracting key regions of training data: Build a spinal image target detection model based on the improved YOLOv8 framework. Perform high-precision recognition and localization of the entire spine to the pelvis in preprocessed images, extracting standardized regional images of key anatomical structures such as the vertebral body, pelvis, and shoulders. Step 3: Build a generative model: Design a latent diffusion model with spatial-channel enhancement (SCELDM) that uses preoperative images as input to generate matching postoperative images. Use internal and external datasets to evaluate the model's image generation and postoperative effect prediction performance.

[0008] Furthermore, in step 1, the collected X-ray image is imported into LabelMe, and the main body is framed using the Rectangle tool. The JSON file generated by LabelMe is loaded, and then the image annotation points in the JSON dictionary are read and converted according to the position of the point relative to the entire image. The relative position and category information of the point are then stored in a txt file as input for the deep learning model.

[0009] Furthermore, in step 2, the network model uses the YOLOv8 deep learning framework to perform target detection on X-ray images. The image processed in step 1 is input into the network model. After training iterations, the network model automatically recognizes the features of the image based on the results of deep learning, detects the image, and intercepts the corresponding full spine and pelvis parts in the X-ray image.

[0010] Furthermore, in step 2, all data sets are processed through the trained YOLOv8 model, and the key areas including the entire spine to the pelvis are intercepted to extract important anatomical structure information.

[0011] Furthermore, in step three, preoperative images are used as input to the SCELDM to generate matching postoperative images. The model training can be divided into two core stages. The first stage, based on the vector quantization generative adversarial network framework, learns a compact discrete latent representation of postoperative X-ray images through quantized encoding, constructing a high-quality feature compression space. The second stage, based on the latent diffusion model, establishes a cross-modal mapping relationship from preoperative image features to postoperative latent representations in the latent space dimension.

[0012] Furthermore, in step three, SCELDM makes key improvements to its conditional guidance mechanism, proposing a spatial-channel enhancement module. The core idea of ​​this module is to enhance features in both the spatial and channel dimensions, and dynamically control the fusion of original and enhanced information through a gating mechanism, effectively improving the model's ability to perceive features in different regions.

[0013] Beneficial effects of the present invention: 1. The YOLOv8 deep learning framework is used to detect objects in X-ray images, automatically locating and capturing the region containing the complete spine and pelvis. By extracting this key anatomical structure as the model's input or region of interest, the model's judgment is not only less susceptible to interference from irrelevant background information, but also guided by a more precise focus on spinal features, thereby improving the fidelity of detail and clinical relevance of the generated results.

[0014] 2. Use the spatial channel enhancement module to perform enhancement in the spatial and channel dimensions, and then use the gating mechanism to dynamically control the fusion of original information and enhanced information to improve the model's ability to perceive the characteristics of different regions.

[0015] 3. Directly predict the postoperative effect, providing a more intuitive and rich reference basis for AIS surgical planning, reducing dependence on subjective experience, and further improving the scientific nature and quality of surgical planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 Spatial channel information enhancement according to an embodiment of the present invention; Figure 3 The adaptive fusion control mechanism of the embodiment of the present invention; Figure 4 Generate an adversarial network for vector quantization according to an embodiment of the present invention; Figure 5 is a potential diffusion model of an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] like Figure 1-Figure 5 As shown, the method for postoperative prediction of scoliosis based on the spatial channel enhanced diffusion model provided by the present invention includes the following steps: 1. Collection, labeling and model training of training data This invention first collects and integrates preoperative and postoperative full-spine X-ray images of patients with adolescent idiopathic scoliosis based on the AIS image database established in the early stage. The data covers different genders, ages and major curve types and is representative.

[0019] Import the image data into the annotation tool LabelMe and use the Rectangle tool to annotate the main body regions, generating a corresponding .json file. Then, read the coordinate information in the json dictionary, convert it to relative positions, and output it in .txt format for training the object detection model.

[0020] Object detection was performed using the YOLOv8 framework, trained on labeled data. Ultimately, a detection model was developed that accurately identifies and crops structures such as the spine and pelvis in X-ray images. Simultaneously, postoperative images were latently encoded using a VQ-GAN to extract their compressed feature representations in latent space. Preoperative images were then paired with postoperative latent features to train a latent diffusion model (SCELDM), establishing a cross-modal mapping from preoperative to postoperative images.

[0021] 2. Target area extraction and structure recognition For example, a 13-year-old female patient with Lenke IIICN AIS had a preoperative Cobb angle of 54° and was scheduled for non-selective orthopedic fusion. A standard standing full-spine X-ray taken preoperatively was imported into the model system.

[0022] First, the YOLOv8 model performs object detection on the preoperative image, automatically identifying and cropping a standardized image encompassing the C1 to pelvic region. The detection results clearly identify key anatomical structures such as vertebral contours, pelvic level, and shoulder position, providing accurate input for the subsequent postoperative prediction model.

[0023] 3. Postoperative Prediction and Inference Process: The patient's preoperative cropped X-ray image was input into the trained SCELDM model. The model first performed a stepwise diffusion sampling within the latent space based on the preoperative information, ultimately generating a matching postoperative X-ray image. This image clearly shows a significant reduction in the major curve after surgery, with a predicted Cobb angle of 16°, left shoulder elevation, and a corrected fusion segment located at T4-L3. Compared to the actual postoperative image, the model-predicted image showed high structural consistency, with a Cobb angle of 19°, an error of 3°, left shoulder elevation, and a corrected fusion segment located at T4-L3, validating the effectiveness of this method.

Claims

1. A method for postoperative scoliosis prediction based on a spatial channel enhanced diffusion model, characterized by: The following steps are involved: Step 1: Training Data Collection, Annotation, and Preprocessing: Preoperative and postoperative X-ray images of adolescent idiopathic scoliosis patients were collected to create a database. The images in the database were annotated and preprocessed to extract the complete spinal-pelvic region from the paired preoperative and postoperative X-ray images, covering the atlanto-occipital joint to the pubic symphysis. Step 2: Extracting key regions of training data: Build a spinal image target detection model based on the improved YOLOv8 framework. Perform high-precision recognition and localization of the entire spine to the pelvis in preprocessed images, extracting standardized regional images of key anatomical structures such as the vertebral body, pelvis, and shoulders. Step 3: Build a generative model: Design a latent diffusion model based on spatial channel enhancement, which uses preoperative images as input and generates matching postoperative images; use internal and external datasets to evaluate the model's image generation and postoperative effect prediction performance.

2. The method for postoperative scoliosis prediction based on the spatial channel enhanced diffusion model according to claim 1, characterized in that: In step 1, the collected X-ray image is imported into LabelMe, and the main body is framed using the Rectangle tool. The JSON file generated by LabelMe is loaded, and then the image annotation points in the JSON dictionary are read and converted according to the position of the point relative to the entire image. The relative position and category information of the point are then stored in a txt file as input to the deep learning model.

3. The method for postoperative scoliosis prediction based on the spatial channel enhanced diffusion model according to claim 2, characterized in that: In step 2, the network model uses the YOLOv8 deep learning framework to perform target detection on the X-ray image. The image processed in step 1 is input into the network model. After training iterations, the network model automatically identifies the features of the image based on the results of deep learning, detects the image, and intercepts the corresponding full spine and pelvis parts in the X-ray image.

4. The method for postoperative scoliosis prediction based on the spatial channel enhanced diffusion model according to claim 3, characterized in that: In step 2, all data sets are processed using the trained YOLOv8 model, and key areas from the entire spine to the pelvis are captured to extract important anatomical structure information.

5. The method for postoperative scoliosis prediction based on the spatial channel enhanced diffusion model according to claim 4, characterized in that: In step three, the preoperative image is used as the SCELDM input to generate the matching postoperative image; the training of this model is divided into two core stages: the first stage is based on the vector quantization generative adversarial network framework, which learns the compact discrete latent representation of the postoperative X-ray image through quantization encoding to construct a high-quality feature compression space; the second stage is based on the latent diffusion model to establish a cross-modal mapping relationship from preoperative image features to postoperative latent representation in the latent space dimension.

6. The method for postoperative scoliosis prediction based on the spatial channel enhanced diffusion model according to claim 5, characterized in that: In the step three, SCELDM proposes a space-channel enhancement module in the conditional guidance mechanism, which enhances the features in the spatial dimension and channel dimension respectively, and dynamically controls the fusion process of the original information and the enhanced information through the gating mechanism.

Images