[论文解读] GasHis-Transformer: A Multi-scale Visual Transformer Approach for Gastric Histopathology Image Classification.

Haoyuan Chen, Chen Li+7 more|arXiv (Cornell University)|Apr 29, 2021

AI in cancer detection参考文献 54被引用 14

一句话总结

该论文提出 GasHis-Transformer，一种用于胃组织病理学图像分类的多尺度视觉变换器模型，整合了全局与局部特征提取模块。在公开的苏木精-伊红染色数据集上，其准确率为98.0%，精确率为98.0%，召回率为100.0%，F1得分为96.0%，展现出高性能及对噪声和对抗性攻击的强鲁棒性。

ABSTRACT

Existing deep learning methods for diagnosis of gastric cancer commonly use convolutional neural network. Recently, the Visual Transformer has attracted great attention because of its performance and efficiency, but its applications are mostly in the field of computer vision. In this paper, a multi-scale visual transformer model, referred to as GasHis-Transformer, is proposed for Gastric Histopathological Image Classification (GHIC), which enables the automatic classification of microscopic gastric images into abnormal and normal cases. The GasHis-Transformer model consists of two key modules: A global information module and a local information module to extract histopathological features effectively. In our experiments, a public hematoxylin and eosin (H&E) stained gastric histopathological dataset with 280 abnormal and normal images are divided into training, validation and test sets by a ratio of 1 : 1 : 2. The GasHis-Transformer model is applied to estimate precision, recall, F1-score and accuracy on the test set of gastric histopathological dataset as 98.0%, 100.0%, 96.0% and 98.0%, respectively. Furthermore, a critical study is conducted to evaluate the robustness of GasHis-Transformer, where ten different noises including four adversarial attack and six conventional image noises are added. In addition, a clinically meaningful study is executed to test the gastrointestinal cancer identification performance of GasHis-Transformer with 620 abnormal images and achieves 96.8% accuracy. Finally, a comparative study is performed to test the generalizability with both H&E and immunohistochemical stained images on a lymphoma image dataset and a breast cancer dataset, producing comparable F1-scores (85.6% and 82.8%) and accuracies (83.9% and 89.4%), respectively. In conclusion, GasHisTransformer demonstrates high classification performance and shows its significant potential in the GHIC task.

研究动机与目标

• 为解决传统卷积神经网络在分类胃组织病理学图像时的局限性，提出基于视觉变换器的方法。
• 通过在多尺度变换器架构中结合全局与局部信息模块，提升胃组织病理学图像的特征提取能力。
• 评估模型在各种图像损坏和对抗性攻击下的鲁棒性，以确保临床可靠性。
• 在不同癌症数据集上测试模型对多种染色类型（苏木精-伊红染色和免疫组织化学染色）的泛化能力。
• 在胃癌、淋巴瘤和乳腺癌数据集上实现与现有方法相当或更优的诊断性能。

提出的方法

• GasHis-Transformer 模型采用双分支架构，包含全局信息模块和局部信息模块，以捕捉长距离与细粒度的组织病理学特征。
• 通过在不同分辨率下处理输入图像块，实现多尺度特征表示，使模型能够学习苏木精-伊红染色胃组织图像中的分层模式。
• 视觉变换器编码器使用自注意力机制建模图像块之间的依赖关系，取代传统的卷积操作。
• 在 [CLS] 标记表示上应用分类头，以预测胃组织状态为正常或异常。
• 通过注入十种噪声类型（包括四种对抗性攻击和六种常规图像损坏）评估鲁棒性。
• 在外部数据集上测试泛化能力，包括使用苏木精-伊红染色和免疫组织化学染色协议的淋巴瘤和乳腺癌图像。

实验结果

研究问题

• RQ1 多尺度视觉变换器模型是否能在胃组织病理学图像分类中超越传统CNN？
• RQ2 全局与局部特征提取的融合在GHIC中如何提升分类性能？
• RQ3 在临床诊断背景下，GasHis-Transformer对图像噪声和对抗性攻击的鲁棒性达到何种程度？
• RQ4 该模型能否在不同染色技术及癌症类型（如淋巴瘤和乳腺癌）间实现良好泛化？
• RQ5 GasHis-Transformer在大规模、具有临床相关性的胃癌数据集上的诊断准确率如何？

主要发现

• 在280张苏木精-伊红染色胃组织病理学图像的测试集中，GasHis-Transformer实现了98.0%的准确率、98.0%的精确率、100.0%的召回率和96.0%的F1得分。
• 该模型表现出强大的鲁棒性，在十种不同噪声类型下（包括对抗性攻击）仍保持高性能。
• 在包含620张异常胃组织图像的更大临床数据集中，该模型在识别胃肠道癌病例方面达到了96.8%的准确率。
• 在淋巴瘤数据集上评估时，GasHis-Transformer在苏木精-伊红染色和免疫组织化学染色下分别取得了85.6%的F1得分和83.9%的准确率。
• 在乳腺癌数据集上，该模型记录了82.8%的F1得分和89.4%的准确率，表明其在不同癌症类型和染色方法间具有出色的泛化能力。