Wang Li, Chen Ruiyun, Weng Jingjing, Li Huiping, Ying Shi, Zhang Jinghui, Yu Zehao, Peng Chengbin, Zheng Siming
Gynaecology Department, Ningbo Medical Centre Lihuili Hospital, Ningbo, Zhejiang, China.
College of Information Science and Engineering, Ningbo University, Ningbo, Zhejiang, China.
Front Oncol. 2024 Nov 22;14:1423782. doi: 10.3389/fonc.2024.1423782. eCollection 2024.
This study aims to investigate the feasibility of employing artificial intelligence models for the detection and localization of cervical lesions by leveraging deep semantic features extracted from colposcopic images.
The study employed a segmentation-based deep learning architecture, utilizing a deep decoding network to integrate prior features and establish a semantic segmentation model capable of distinguishing normal and pathological changes. A two-stage decision model is proposed for deep semantic feature mining, which combines image segmentation and classification to categorize pathological changes present in the dataset. Furthermore, transfer learning was employed to create a feature extractor tailored to colposcopic imagery. Multi-scale data were bolstered by an attention mechanism to facilitate precise segmentation of lesion areas. The segmentation results were then coherently mapped back onto the original images, ensuring an integrated visualization of the findings.
Experimental findings demonstrated that compared to algorithms solely based on image segmentation or classification, the proposed approach exhibited superior accuracy in distinguishing between normal and lesioned colposcopic images. Furthermore, it successfully implemented a fully automated pixel-based cervical lesion segmentation model, accurately delineating regions of suspicious lesions. The model achieved high sensitivity (96.38%), specificity (95.84%), precision (97.56%), and f1 score (96.96%), respectively. Notably, it accurately estimated lesion areas, providing valuable guidance to assisting physicians in lesion classification and localization judgment.
The proposed approach demonstrates promising capabilities in identifying normal and cervical lesions, particularly excelling in lesion area segmentation. Its accuracy in guiding biopsy site selection and subsequent localization treatment is satisfactory, offering valuable support to healthcare professionals in disease assessment and management.
本研究旨在探讨利用从阴道镜图像中提取的深度语义特征，运用人工智能模型检测和定位宫颈病变的可行性。
该研究采用基于分割的深度学习架构，利用深度解码网络整合先验特征，建立能够区分正常和病理变化的语义分割模型。提出了一种用于深度语义特征挖掘的两阶段决策模型，该模型结合图像分割和分类对数据集中存在的病理变化进行分类。此外，采用迁移学习创建了一个针对阴道镜图像定制的特征提取器。通过注意力机制增强多尺度数据，以促进病变区域的精确分割。然后将分割结果连贯地映射回原始图像，确保结果的综合可视化。
实验结果表明，与仅基于图像分割或分类的算法相比，所提出的方法在区分正常和病变阴道镜图像方面具有更高的准确性。此外，它成功实现了一个基于像素的全自动宫颈病变分割模型，准确地勾勒出可疑病变区域。该模型分别达到了高灵敏度（96.38%）、特异性（95.84%）、精度（97.56%）和f1分数（96.96%）。值得注意的是，它准确地估计了病变面积，为协助医生进行病变分类和定位判断提供了有价值的指导。
所提出的方法在识别正常和宫颈病变方面显示出有前景的能力，尤其在病变区域分割方面表现出色。其在指导活检部位选择和后续定位治疗方面的准确性令人满意，为医疗专业人员在疾病评估和管理中提供了有价值的支持。
