Research On Spinal CT Image Segmentation Method Based On Deep Learning

Spinal diseases pose a significant threat to patients’ lives and overall well-being.Owing to modern living and work patterns,these diseases are exhibiting a growing trend in incidence rates,along with a younger age of onset.In the clinical diagnosis and management of spinal diseases,CT imaging holds a pivotal position that cannot be substituted.With advancements in medical image processing and computer-aided diagnosis,automatic segmentation algorithms for spinal CT images have come into the forefront of clinical applications.These algorithms effectively aid doctors in precisely and promptly pinpointing abnormalities or lesions,thereby alleviating their workload during the diagnostic process and enhancing both diagnostic efficiency and accuracy.Consequently,automatic segmentation algorithms for spinal CT images have garnered increasing attention and research interest within the realm of medical image processing.Currently,deep learning-based segmentation algorithms for spinal CT images significantly outperform other types of algorithms in terms of accuracy,thereby establishing them as the preferred approach in this field.Depending on the nature of the segmentation target,research in this domain can be categorized into two primary directions: semantic segmentation,which focuses on the overall spine,and instance segmentation,which emphasizes the discrimination of individual vertebral bones.This article endeavors to delve into the common challenges faced by both directions,as well as the interface between these two types of segmentation.These challenges primarily include:inconsistencies in image resolution and size arising from different acquisition devices;variations in vertebral bone size across different scanning fields;and indistinct boundaries between vertebrae and adjacent tissues.These issues have a profound impact on the precision of spinal image segmentation.Notably,while these two types of segmentation are frequently studied independently,their synergistic use is imperative in practical applications.Current research on instance segmentation predominantly relies on the cascading of multiple deep neural networks.However,this approach often leads to longer training times,an increase in the total number of network parameters,and a subsequent surge in memory and computational requirements.Alternatively,some research efforts aim to supplant semantic segmentation with iterative instance segmentation.While this approach circumvents the multistage issue,it still struggles with missed detections and excessive memory consumption,as it necessitates the detection and storage of each individual vertebral bone during the training process.Furthermore,as deep learning-based methods,they often encounter difficulties due to their huge parameter counts and high complexity,rendering deployment a challenging task.Addressing the challenges associated with spinal semantic segmentation’s limited accuracy,this article introduces the following innovative advancements:1.We present a novel U-Net architecture,enhanced with residual multi-scale fusion and attention skip connections.This innovative network incorporates residual feature pyramid blocks and attention skip structures.The residual feature pyramid blocks excel at capturing multi-scale features of vertebrae,accommodating various shapes and sizes,and seamlessly fusing these extracted features to merge global and local information.Meanwhile,the attention skip structure is specifically designed to harmoniously integrate shallow and deep features,effectively mitigating the disruptive influence of redundant shallow features on the segmentation process.This structure also facilitates the downward flow of information from the decoding section.Furthermore,to optimize the segmentation outcomes,we utilize a joint loss function.At each decoding layer,the features are meticulously upsampled to match the original image size,allowing the network to establish a multi-level segmentation mechanism that significantly enhances boundary segmentation clarity.Extensive experiments conducted on the Ver Se2019 and Ver Se2020 datasets validate the effectiveness of our proposed method.Notably,in terms of key evaluation metrics such as Dice and HD,our approach demonstrates impressive performance,surpassing most state-of-the-art models.2.We introduce a framework called CDLF specifically designed for instance segmentation of spinal CT images,aiming to tackle the challenge of low accuracy in this domain.This framework integrates semantic segmentation with the Mask RCNN instance segmentation framework through wavelet fusion,significantly enhancing the segmentation accuracy of Mask RCNN.Central to our approach is the novel OSU-Net architecture,which serves as the semantic segmentation component,providing crucial overall localization information for instance segmentation.To further boost the performance of semantic segmentation,we introduce the OSAv2 module within the encoder section of OSU-Net and propose an innovative upsampling attention module for the decoder section.The integration of these modules ensures that our framework can capture and utilize key features effectively.Experimental evaluations conducted on the Ver Se2019 dataset reveal that our proposed method achieves exceptional segmentation performance in spinal instance segmentation.When compared to other state-of-the-art spinal segmentation methods,our framework demonstrates competitive and often superior results,highlighting its potential in addressing the challenges of instance segmentation in spinal CT images.3.We introduce HTNet-DSFP,an end-to-end,single-stage vertebral instance segmentation network,designed to overcome the challenges posed by training difficulties and time-consuming iterations associated with traditional multiple deep neural network cascades or iterative instance segmentation methods.At the backbone of HTNet-DSFP lies Trans-Res,a novel architecture that seamlessly integrates convolutional layers with a Transformer using residual connections.This integration allows our network to capture intricate local details while maintaining a focus on long-term dependencies.Moreover,we propose a Feature Pyramid for Dynamic Selection which effectively fuses and enhances contextual information across various scales.This ensures that our network can accurately segment vertebrae,regardless of their size or location within the image.Our method is thoroughly evaluated on the Ver Se2019 and Ver Se2020 datasets,demonstrating its superior performance in vertebral segmentation tasks.By harnessing the complementary strengths of convolutional and Transformer-based architectures,HTNet-DSFP offers a robust and efficient solution that surpasses the limitations of previous approaches.4.We introduce a lightweight vertebral instance segmentation network that cleverly integrates CNN and Transformer technologies to address the challenges of excessive parameter counts and memory consumption in traditional deep learning methods.This innovative approach facilitates the deployment of vertebral segmentation models on edge devices,enabling real-time or near-real-time analysis in resource-constrained environments.To strike a balance between segmentation accuracy and model complexity,we introduce a lightweight dual residual depth-wise convolutional block in the initial stages of our network.This block efficiently extracts local details by leveraging depth-wise separable convolutions,significantly reducing the overall parameter count.In the subsequent stages,we seamlessly integrate CNN and Transformer capabilities to design a dual-path attention downsampling module.This module effectively captures contextual information,enhancing the model’s ability to identify vertebral boundaries and structures.Extensive comparative experiments on the Ver Se2019 dataset demonstrate the superiority of our approach.Our model outperforms other backbone networks in terms of segmentation accuracy while maintaining a significantly lower parameter count.This balance between performance and efficiency positions our model as a viable solution for vertebral instance segmentation in resource-constrained scenarios,such as edge computing.