乳腺钼靶肿块自监督预训练迁移检测方法研究

PDF下载HTML

Title:

Self-supervised pretraining detection of mammographic mass targets in breast

作者:

莫宏伟¹, 孙琪¹, 孙鹏¹, 张显玉², 孙江宏³, 孙惟嘉³

1. 哈尔滨工程大学 智能科学与工程学院, 黑龙江 哈尔滨 150001;
2. 哈尔滨医科大学附属肿瘤医院 乳腺外科, 黑龙江 哈尔滨 150001;
3. 哈尔滨医科大学附属肿瘤医院 影像中心, 黑龙江 哈尔滨 150001

Author(s):

MO Hongwei¹, SUN Qi¹, SUN Peng¹, ZHANG Xianyu², SUN Jianghong³, SUN Weijia³

1. College of Intelligent Systems Science and Engineering, Harbin Engineering University, Harbin 150001, China;
2. Breast Surgery Department, Cancer Hospital Affiliated to Harbin Medical University, Harbin 150001, China;
3. Imaging Center, Cancer Hospital Affiliated to Harbin Medical University, Harbin 150001, China

关键词:

目标检测; 自监督; 钼靶影像; 预训练; 数据增强; 视觉表示; 卷积神经网络; 图像分类

Keywords:

object detection; self-supervised; mammographic; pretraining; data augment; visual representation; convolution neural network; image classification

分类号:

TP181

DOI:

10.11992/tis.202304032

文献标志码:

2024-09-02

摘要:

借助深度学习技术在乳腺钼靶领域辅助医生进行乳腺癌诊断在当下已经成为很多研究关注的热点，诊断技术主要包括良恶性分类、病灶区域检测以及病灶区域分割等。由于深度学习训练的模型性能很大程度上依赖于大量的带有标注的数据，而医学图像数据集往往存在数据量少、标注成本昂贵以及公开数据集标注质量差等现象，所以在医学图像领域应用深度学习技术具有重重困难。为使基于深度学习的乳腺钼靶计算机辅助诊断技术的开发不受限于大量有标注的数据，提出一种适用于钼靶自监督目标检测方法来完成乳腺钼靶肿块检测任务，利用大量来自肿瘤医院的数据预训练，并在公开数据集DDSM上进行微调与测试。实验结果表明，提出模型在乳腺钼靶肿块检测任务中表现优异，并且不依赖于位置标签，具有重要的研究价值与应用前景。

Abstract:

Deep learning technology has become a major research focus in assisting doctors with breast cancer diagnosis, particularly in mammographic imaging. This research encompasses tasks such as benign and malignant classification, lesion region detection, and lesion region segmentation. However, the performance of deep learning models heavily relies on large amounts of labeled data. Medical image datasets often suffer from limited data, high labeling costs, and poor labeling quality of open datasets, which make it challenging to apply deep learning technology effectively in the medical imaging field. To advance computer-aided diagnosis of breast mammographic targets using deep learning without relying on large amounts of labeled data, this study proposes a method for self-supervised target detection in mammographic images. The method addresses the task of detecting breast mammographic masses. Extensive data from cancer hospitals were pretrained, fine-tuned, and tested on the open dataset DDSM. The experimental results show that the proposed model performs excellently in detecting breast mammographic masses and does not rely on location tags. These results highlight the significant research value of the model and promising application prospects.