| 安政,韩乐,史明,周云飞,张珈豪.基于双流判别器的生成对抗皮肤病变分割网络[J].中国医学影像技术,2024,40(12):1914~1919 |
| 基于双流判别器的生成对抗皮肤病变分割网络 |
| Skin lesion segmentation network with dual-stream discriminator based on generative adversarial networks |
| 投稿时间：2024-07-30 修订日期：2024-12-10 |
| DOI：10.13929/j.issn.1003-3289.2024.12.023 |
| 中文关键词: 皮肤病 神经网络,计算机 皮肤镜检查 |
| 英文关键词:skin diseases neural networks, computer dermoscopy |
| 基金项目: |
|
| 摘要点击次数: 318 |
| 全文下载次数: 100 |
| 中文摘要: |
| 目的 观察超像素引导含双流基于局部图像判别器生成对抗网络(SPDD-PatchGAN)用于分割皮肤镜图像中皮肤病变的价值。方法 分别于国际皮肤成像合作组织(ISIC)2016数据集及Human Against Machine with 10 000 training images(HAM10000)数据集中收集1 279及10 015幅皮肤病变患者的皮肤镜图像,以数据集中的手动分割结果为参考标准,利用包含多尺度上下文提取模块(MCEM)的残差注意力UNet(RA-UNet)为生成器,采用基于局部图像的超像素引导的双流判别策略为判别器构建SPDD-PatchGAN并以之分割图像中的皮肤病变;与深度卷积生成对抗网络(DCGAN)、UNet、Attention-UNet、上下文编码器网络(CENet)、上下文金字塔融合网络(CPFNet)和生成对抗模型双流生成对抗网络(DAGAN)对比,采用平均交并比(mIoU)、准确率(Accuracy)、召回率(Recall)评价SPDD-PatchGAN的分割效能。结果 SPDD-PatchGAN分割皮肤病变的整体效果较佳,且其mIoU、Accuracy及Recall均优于DCGAN、UNet、Attention-UNet、CENet、CPFNet及DAGAN。结论 SPDD-PatchGAN可有效分割皮肤镜图像中的皮肤病变。 |
| 英文摘要: |
| Objective To observe the value of superpixel-guided generative adversarial network with dual-stream patch-based discriminator (SPDD-PatchGAN) for segmenting skin lesions in dermatoscopy images. Methods A total of 1 279 and 10 015 dermatoscopic images of patients with skin lesions were collected from International Skin Imaging Collaboration (ISIC) 2016 and Human Against Machine with 10 000 training images (HAM10000) datasets, respectively. Taken manual segmentation results as reference standards, residual attention UNet (RA-UNet) with multi-scale context extraction module (MCEM) as generator and dual stream discrimination strategy guided by superpixels based on local images as the discriminator, SPDD-PatchGAN was constructed to segment skin lesions in dermatoscopy images, and the results were compared with those of deep convolutional generative adversarial network (DCGAN), UNet, Attention-UNet, context encoder network (CENet), context pyramid fusion network (CPFNet) and generative adversarial network with dual discriminator (DAGAN). The segmenting performance of SPDD-PatchGAN was evaluated using the mean intersection over union (mIoU), Accuracy and Recall. Results The overall effect of SPDD-PatchGAN for segmenting skin lesions in dermatoscopy images was better, with mIoU, Accuracy and Recall superior to DCGAN, UNet, Attention-UNet, CENet, CPFNet and DAGAN. Conclusion SPDD-PatchGAN could effectively segment skin lesions in dermatoscopy images. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
