混合CNN与ViT模型，在小数据下实现高召回率肺炎检测。
CNN-ViT Hybrid for Pneumonia Detection: Theory and Empiric on Limited Data without Pretraining
- 结合CNN局部特征提取与ViT全局建模优势，设计新混合架构。
- 在50%数据下达到0.9443的最高召回率，F1分数稳定在0.85左右。
- 适用于数据量小且类别不平衡的医学图像诊断场景。
本研究探索了在数据量有限且存在明显类别不平衡的情况下，将卷积神经网络（CNN）与视觉变换器（ViT）进行混合建模的可行性。训练从零开始，专注于理论分析与实验验证所提混合模型的架构优势。在不同数据比例下，对比平衡与非平衡数据集的实验结果表明：该混合模型通过融合CNN与ViT的优点，在50%数据量的平衡数据集中实现了0.9443的最高召回率，且在所有数据比例下F1分数维持在约0.85，表现出良好的诊断一致性；同时在类别不平衡数据中表现优于单独的CNN与ViT。尽管模型结构更复杂，但在所有数据比例下的训练时间与纯变压器模型相当。
原文摘要 · Abstract (English)
This research explored the hybridization of CNN and ViT within a training dataset of limited size, and introduced a distinct class imbalance. The training was made from scratch with a mere focus on theoretically and experimentally exploring the architectural strengths of the proposed hybrid model. Experiments were conducted across varied data fractions with balanced and imbalanced training datasets. Comparatively, the hybrid model, complementing the strengths of CNN and ViT, achieved the highest recall of 0.9443 (50% data fraction in balanced) and consistency in F1 score around 0.85, suggesting reliability in diagnosis. Additionally, the model was successful in outperforming CNN and ViT in imbalanced datasets. Despite its complex architecture, it required comparable training time to the transformers in all data fractions.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。