%0 Journal Article %T SMLS：一种用于宫颈图像识别的分类识别模型
SMLS: A Classification Recognition Model for Cervical Image Recognition %A 方庆斌 %A 邹任玲 %A 许婧 %A 管睿 %A 谷雪莲 %A 胡秀芳 %A 尹学志 %J Modeling and Simulation %P 83-96 %@ 2324-870X %D 2025 %I Hans Publishing %R 10.12677/mos.2025.143205 %X 宫颈癌是全球妇女中发病率第四高、最危险的癌症。不过，只要及时发现和治疗，治愈率几乎可以达到100%。阴道镜检查是一种常见的医疗程序，用于评估宫颈病变的严重程度。然而，随着宫颈癌病例数量的逐年上升，医生面临的工作量也在不断增加，导致人工视觉检查中可能出现诊断错误和漏诊。由于传统ViT模型的特征提取能力并不能充分利用且运行效率低，因此本研究提出了一种新的模型，通过深度可分卷积和多层次多尺度特征融合(deeply separable convolution and multi-level multi-scale feature fusion, SMLS)对ViT模型进行改进，并辅以迁移学习以提高准确性。该方法旨在将宫颈阴道镜图像分为正常、宫颈上皮内瘤变和浸润癌三个类别。首先对数据集进行数据扩充以扩大其规模，然后对宫颈图像数据集的模型参数进行微调，然后通过与四种传统神经网络模型进行对比验证该模型的有效性，最后利用消融实验来进行对模型的各个模块的有效性验证。实验结果表明，所提出的方法在有限的数据集上达到了87.80%的准确率。该方法能够充分利用该模型的特征提取能力，识别效果好，可以应用于对宫颈图像的快速诊断识别。
Cervical cancer is the fourth most prevalent and dangerous female cancer in the world, but if detected and treated in a timely manner, the cure rate is almost 100%. Colposcopy is a common medical procedure used to assess the severity of cervical lesions. However, with the number of cervical cancer cases rising every year, the workload faced by physicians is increasing, leading to potential diagnostic errors and missed diagnoses during manual visualization. Since the feature extraction capability of the traditional Vision Transformer (ViT) model is not fully utilized and operates inefficiently, this study introduces a novel model, which improves the ViT model by using deeply separable convolution and multi-level multi-scale feature fusion (SMLS), and is used to perform a new method of recognition and classification of cervical images, supplemented with migration learning to improve the accuracy. The objective of the method is to categorize cervical colposcopy images into three classifications: normal, cervical intraepithelial neoplasia and invasive carcinoma. Firstly, data expansion is performed on the dataset to enlarge its size, then the model parameters of the cervical image dataset are fine-tuned, and then the efficacy of the model is validated through comparing it with four traditional neural network models, and finally ablation experiments are used to carry out the validation of the efficacy of each module of the model. The experimental findings demonstrate that the proposed method achieves an accuracy of 87.80% on a limited dataset. The method can make full use of the feature extraction ability of the model, with good recognition effect, and can be applied to the rapid diagnosis and recognition of cervical images. %K 宫颈癌， %K SMLS模型， %K 宫颈图像识别， %K 多层次多尺度特征融合， %K 可分离卷积
Cervix %K SMLS Model %K Cervical Image Recognition %K Multi-Level and Multi-Scale Feature Fusion %K Separable Convolution %U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=109032