跳转至

文章背景与核心概要

肺癌仍然是全球癌症相关死亡的主要原因。尽管深度学习为计算病理学提供了极具前景的途径，但将组织分类和区域分割有效整合到一个统一分析框架中的全面基准测试依然十分稀缺。

本文介绍了一种专为多类组织分类和像素级组织病理学区域分割而设计的两阶段深度学习框架。通过对每个阶段的最先进架构进行系统性评估，该研究为自动组织病理学图像分析建立了一个准确、计算高效且可复现的基准。

肺癌组织病理学深度学习架构的综合基准测试 (Comprehensive Benchmarking of Deep Learning Architectures for Lung Cancer Histopathology)

arXiv ID: 2608.15915
主要学科: 计算机视觉与模式识别 (cs.CV)
次要学科: 人工智能 (cs.AI)
提交日期: 2026年8月16日
作者: Hadi Hasan, Safaa Salman, Lama Sleem, Ralph Mouawad, Ali Chehab

arXiv ID: 2608.15915
Primary Subject: Computer Vision and Pattern Recognition (cs.CV)
Secondary Subjects: Artificial Intelligence (cs.AI)
Submitted On: August 16, 2026
Authors: Hadi Hasan, Safaa Salman, Lama Sleem, Ralph Mouawad, Ali Chehab

📝 摘要

肺癌仍然是全球癌症相关死亡的主要原因。尽管深度学习为计算病理学提供了极具前景的途径，但将组织分类和区域分割有效整合到一个统一分析框架中的全面基准测试依然十分稀缺。

本文介绍了一种专为多类组织分类和像素级组织病理学区域分割而设计的两阶段深度学习框架。通过对每个阶段的最先进架构进行系统性评估，该研究为自动组织病理学图像分析建立了一个准确、计算高效且可复现的基准。

📝 Summary

Lung cancer remains the leading cause of cancer-related mortality globally. While deep learning offers promising avenues for computational pathology, comprehensive benchmarks that effectively integrate tissue classification and region segmentation into a unified analytical framework are scarce.

This paper introduces a two-stage deep learning framework designed for multi-class tissue classification and pixel-level histopathological region segmentation. By systematically evaluating state-of-the-art architectures at each stage, the study establishes an accurate, computationally efficient, and reproducible baseline for automated histopathological image analysis.

🔬 方法论与架构

1. 组织分类阶段

研究人员在结合了 LC25000 和 LungHist700 数据集的共 39,000 张图像上评估了六种不同的模型。其目标是区分以下三个类别： * 腺癌 (Adenocarcinoma) * 鳞状细胞癌 (Squamous cell carcinoma) * 正常肺组织 (Normal lung tissue)

评估的模型: * 自定义卷积神经网络 (CNN) * VGG16 * DenseNet * MobileNetV3 * 自定义视觉Transformer (ViT) * YOLO11

🔬 Methodology & Architecture

1. Tissue Classification Stage

The researchers evaluated six distinct models on a combined dataset of 39,000 images sourced from LC25000 and LungHist700. The objective was to differentiate between three categories: * Adenocarcinoma * Squamous cell carcinoma * Normal lung tissue

Models Evaluated: * Custom Convolutional Neural Network (CNN) * VGG16 * DenseNet * MobileNetV3 * Custom Vision Transformer (ViT) * YOLO11

2. 区域分割阶段

对于像素级区域分割，研究使用 GlaS 腺体分割基准 对四种架构进行了基准测试： * U-Net * 骨干网络为 ResNet 的 U-Net (ResNet-encoder U-Net) * DeepLabV3+ * YOLO11-seg

2. Region Segmentation Stage

For pixel-level region segmentation, four architectures were benchmarked using the GlaS gland segmentation benchmark: * U-Net * ResNet-encoder U-Net * DeepLabV3+ * YOLO11-seg

3. 端到端集成

随后，将两个阶段中表现最好的模型耦合到一个统一的端到端框架中。

3. End-to-End Integration

The highest-performing models from both stages were subsequently coupled into a unified, end-to-end framework.

📊 核心结果

  • 组织分类性能：
  • YOLO11 实现了最高的整体分类性能。
  • 准确率 (Accuracy): 98.38%
  • 5折交叉验证准确率 (5-Fold Cross-Validation Accuracy): 98.21 ± 0.35%

  • 宏平均 F1 分数 (Macro F1-Score): 0.98

  • 区域分割性能：

  • DeepLabV3+ 获得了最高的准确率指标，交并比 (IoU) 为 0.80，Dice 分数为 0.89。
  • YOLO11-seg 实现了极具竞争力的 0.79 的 IoU，同时其参数量大约减少了 14 倍，展现出卓越的计算效率。

📊 Key Results

  • Tissue Classification Performance:
  • YOLO11 achieved the highest overall classification performance.
  • Accuracy: 98.38%
  • 5-Fold Cross-Validation Accuracy: 98.21 ± 0.35%

  • Macro F1-Score: 0.98

  • Region Segmentation Performance:

  • DeepLabV3+ attained the highest accuracy metrics with an Intersection over Union (IoU) of 0.80 and a Dice score of 0.89.
  • YOLO11-seg achieved a highly competitive IoU of 0.79 while utilizing approximately 14x fewer parameters, demonstrating superior computational efficiency.

🔗 链接与资源