[email protected]

国际临床研究杂志

International Journal of Clinical Research

您当前位置：首页 > 精选文章

International Journal of Clinical Research. 2026; 10: (9) ; 10.12208/j.ijcr.20260441 .

Research on lung cancer CT classification and lesion segmentation based on transfer learning
肺癌CT影像分类与病灶分割方法研究

作者： 张钧浩 *, 冯泓翔

大连东软信息学院健康医疗科技学院 辽宁大连

*通讯作者： 张钧浩,单位：大连东软信息学院健康医疗科技学院 辽宁大连; ；

引用本文: 张钧浩, 冯泓翔 肺癌CT影像分类与病灶分割方法研究[J]. 国际临床研究杂志, 2026; 10: (9) : 76-79.
Published: 2026/9/10 10:55:28

摘要

以IQ-OTH/NCCD肺部CT数据集（1190张）为对象，采用限制对比度自适应直方图均衡化（CLAHE）增强、高斯滤波、非锐化掩模（USM）锐化预处理提升病灶特征；利用ImageNet预训练视觉几何组 16 层网络（VGG16）两阶段迁移学习实现三分类；传统图像处理自动生成伪标签掩膜监督U-Net实现像素级病灶分割。实验表明：预处理VGG16准确率98.74%，优于ResNet50（97.48%）和EfficientNet-B0（97.06%）；U-Net分割像素准确率99.29%、IoU 65.3%、Dice 78.6%。方法无需手工像素标注，实现分类与分割联合诊断。

关键词： 肺部CT影像；VGG16；迁移学习；U-Net；语义分割

Abstract

Taking the IQ-OTH/NCCD lung CT dataset (1190 images) as object, Contrast Limited Adaptive Histogram Equalization (CLAHE) enhancement, Gaussian filtering and Unsharp Mask (USM) sharpening preprocessing are used to improve lesion features. Two-stage transfer learning based on ImageNet pretrained Visual Geometry Group 16 layer network (VGG16) realizes three-category classification. Pseudo-label masks generated by traditional image processing supervise U-Net for pixel-level lesion segmentation. Experiments show VGG16 accuracy reaches 98.74%, superior to ResNet50 (97.48%) and EfficientNet-B0 (97.06%). U-Net achieves pixel accuracy 99.29%, IoU 65.3% and Dice 78.6%. This method realizes joint diagnosis without manual pixel annotation.

Key words： Lung CT image; VGG16; Transfer learning; U-Net; Semantic segmentation

参考文献 References

[1] 国家癌症中心. 2024年全国癌症统计报告[R]. 北京: 国家癌症中心, 2024.

[2] SHEN W, ZHOU M, YANG W, et al. Multi-crop convolutional neural networks for lung nodule classification[J]. Pattern Recognition, 2016, 61: 630-639.

[3] ARDILA D, KIRALY A P, BHARADWAJ S, et al. End-to-end lung cancer screening with three-dimensional deep learning[J]. Nature Medicine, 2019, 25(6): 954-961.

[4] RONNEBERGER O, FISCHER P, BROX T. U-Net: Convolutional Networks for Biomedical Image Segmentation[M]. Cham: Springer, 2015.

[5] SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[C]. San Diego: ICLR, 2015.

[6] HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]. Las Vegas: IEEE, 2016.

[7] Dominyka Stragyte, Gvidas Mikalauskas, Katrina Gaidulevic, Renata Paukstaitiene, Kestutis Stasaitis,Vidas Raudonis,Skaidra Valiukeviciene.A Structured Review of Deep Learning Approaches and Image-Preprocessing Techniques for Automated Contact Allergy Patch Test Interpretation.[J].Medical sciences (Basel, Switzerland), 2026, 14 (2):322-322.

[8] TIAN Kaiyuan, QIAO Linbo, LIU Baihui, JIANG Gongqingjian, LI Shanshan, LI Dongsheng. A survey on memory-efficient transformer-based model training in AI for science[J].Frontiers of Computer Science,2026,20 (11):