Computer-aided diagnosis system of lung carcinoma using Convolutional Neural Networks

Fangjian Han, Li Yu, Yi JiangWorkshops @ CVPR 2020 · DOI

Introduction

The global incidence of lung cancer continues to rise, placing an immense burden on pathology departments. Manual analysis of hematoxylin and eosin (H&E) stained slides is labor-intensive and susceptible to human error. While digital pathology and artificial intelligence (AI) offer promising solutions, the sheer size of "Whole Slide Images" (WSIs)—often exceeding 10 billion pixels—presents significant computational challenges. Most current research focuses on high-magnification images (20x or 40x), which require substantial memory and processing power, hindering real-time clinical deployment.

This paper presents a lightweight, high-efficiency computer-aided diagnosis (CAD) system using Deep Convolutional Neural Networks (DCNNs) to identify lung carcinoma. By optimizing for 10x magnification and utilizing the ResNet-18 architecture, the system achieves high diagnostic accuracy while minimizing computational overhead for "on-the-edge" clinical use.

Methods

The proposed system follows a streamlined pipeline from digitization to classification.

Overview of the CAD system pipeline, including digitization, segmentation, randomization, and ResNet-based training.

Model Architecture

The researchers utilized a ResNet-18 architecture, selected for its optimal balance between depth and inference speed. ResNet-18 employs residual connections to mitigate the vanishing gradient problem, allowing for effective training of deep networks.

Detailed architecture of the ResNet model used for feature extraction and classification.

Magnification and Tiling

A key innovation is the evaluation of magnification scaling. The team compared performance at 10x, 20x, and 40x magnifications. To process the massive WSIs, a self-developed scanner automatically segments images into $256 \times 256$ pixel tiles, which are subsequently cropped to $224 \times 224$ for DCNN input.

Hardware and Optimization

While training utilized high-end NVIDIA RTX 2080Ti GPUs, the final model was deployed on an ARM Cortex-A57 (NVIDIA Jetson TX2). This demonstrates the system's capability to run on portable, embedded hardware with limited RAM (8GB), facilitating integration directly into clinical scanners.

Datasets

The study leveraged a robust dataset of approximately 5,000 slides:

  • TCGA (The Cancer Genome Atlas): 500 slides served as a baseline.
  • Clinical Data: 4,500 H&E stained lung tissue slides were collected from Xiangya Hospital and other regional centers over two years.
  • Composition: The final dataset comprised 1,650 normal tissue sections and 3,350 lung cancer sections, partitioned into training (70%), validation (15%), and testing (15%) sets.

Results and Discussion

The system demonstrated that high-resolution imagery is not always a prerequisite for accurate diagnosis.

  • Diagnostic Performance: The ResNet-18 model achieved an AUC of 0.976 at 10x magnification. While 40x magnification reached an AUC of 0.999, the 10x resolution provided a sufficient balance for screening.
  • Efficiency: Using 10x images reduces memory consumption to 1/16th of that required for 40x images. This efficiency allows a standard $15 \times 15\text{mm}$ slide to be digitized and analyzed in under 60 seconds.
  • Model Comparison: ResNet-18 performed comparably to the deeper ResNet-50 but required half the training and inference time, reinforcing its suitability for real-time applications.

Conclusion

This research proves that AI-driven diagnostics can be both accurate and lightweight. By proving that 10x magnification is "good enough" for tumor identification, the authors pave the way for integrating AI directly into digital scanners at the clinical edge. Future work will focus on multi-class classification for lung cancer subtyping and ensuring the detection of minute micro-metastases at lower resolutions.

Summary updated

BibTeX
@inproceedings{han2020computer,
  author = {Fangjian Han and Li Yu and Yi Jiang},
  title = {{Computer-aided diagnosis system of lung carcinoma using Convolutional Neural Networks}},
  booktitle = {Workshops @ CVPR},
  year = {2020},
  doi = {10.1109/cvprw50498.2020.00353},
}
Figure