[论文解读] Efficient Lung Cancer Image Classification and Segmentation Algorithm Based on Improved Swin Transformer

Ruina Sun, Yuexin Pang|arXiv (Cornell University)|Jul 4, 2022

Radiomics and Machine Learning in Medical Imaging被引用 5

一句话总结

本文提出了一种改进的Swin Transformer架构，用于高效肺部癌症图像分类与分割，通过自注意力机制和层次化特征学习提升准确率与计算效率。该方法在使用预训练的情况下，分类准确率最高达82.3%，分割准确率超过95%，在真实临床诊断场景中表现出色。

ABSTRACT

With the development of computer technology, various models have emerged in artificial intelligence. The transformer model has been applied to the field of computer vision (CV) after its success in natural language processing (NLP). Radiologists continue to face multiple challenges in today's rapidly evolving medical field, such as increased workload and increased diagnostic demands. Although there are some conventional methods for lung cancer detection before, their accuracy still needs to be improved, especially in realistic diagnostic scenarios. This paper creatively proposes a segmentation method based on efficient transformer and applies it to medical image analysis. The algorithm completes the task of lung cancer classification and segmentation by analyzing lung cancer data, and aims to provide efficient technical support for medical staff. In addition, we evaluated and compared the results in various aspects. For the classification mission, the max accuracy of Swin-T by regular training and Swin-B in two resolutions by pre-training can be up to 82.3%. For the segmentation mission, we use pre-training to help the model improve the accuracy of our experiments. The accuracy of the three models reaches over 95%. The experiments demonstrate that the algorithm can be well applied to lung cancer classification and segmentation missions.

研究动机与目标

• 为缓解放射科医生日益增长的诊断负担，开发一种自动化、高准确率的肺部癌症检测系统。
• 在肺部癌症影像中，超越传统方法，进一步提升分类与分割的准确率。
• 通过改进Swin Transformer的网络架构，提升模型效率与性能。
• 在真实世界医学影像数据上验证该方法，并通过多种指标进行稳健评估。

提出的方法

• 设计了一种改进的Swin Transformer架构，采用移位窗口自注意力机制，在保持计算效率的同时捕捉长距离依赖关系。
• 采用层次化特征提取方法，从肺部CT扫描中捕获多尺度表征。
• 通过大规模医学图像数据集的预训练实现迁移学习，提升模型泛化能力与性能。
• 使用交叉熵损失与Dice损失，对模型进行端到端微调，以同时完成分类与实例分割任务。
• 应用数据增强与归一化技术，提升模型鲁棒性并减少过拟合。
• 该框架支持联合分类与分割，实现临床决策支持的统一推理。

实验结果

研究问题

• RQ1 改进的Swin Transformer架构是否能在肺部癌症分类中实现高于标准模型的准确率？
• RQ2 预训练在肺部CT影像中对基于Transformer的模型分割性能有何影响？
• RQ3 所提出方法在保持高诊断准确率的同时，能在多大程度上降低计算成本？
• RQ4 该模型能否在不同真实世界肺部癌症影像数据集中实现良好泛化？

主要发现

• 所提方法在使用Swin-T进行常规训练以及Swin-B在双分辨率下进行预训练时，最高分类准确率达到82.3%。
• 在预训练条件下，所有三种模型（Swin-T、Swin-S、Swin-B）的肺部癌症分割准确率均超过95%。
• 预训练显著提升了分割性能，表明其在医学影像中常见低数据场景下的重要价值。
• 该模型在不同图像分辨率与数据分布下均展现出强大的泛化能力与鲁棒性。