# Single-view 2D CNNs with Fully Automatic Non-nodule Categorization for False Positive Reduction in Pulmonary Nodule Detection

**Authors:** Hyunjun Eun, Daeyeong Kim, Chanho Jung, Changick Kim  
**arXiv:** [2003.04454](https://arxiv.org/abs/2003.04454)  
**Date:** 2020-03-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.04454v1/

## Key points

- Study focuses on improving accuracy of pulmonary nodule detection by reducing false positives
- Researchers propose a novel framework using ensemble of 2D CNNs with single views
- Imbalance between nodules and non-nodules is a challenge in reducing false positives
- Ensemble of 2D CNNs with single-view 2D patches overcomes problems of computational complexity and intra-class diversity
- Non-nodules are categorized automatically using autoencoder and k-means clustering, eliminating manual categorization by radiologists
- Training scheme extends learning capability and resolves difficulties in extracting features from non-nodules with large appearance variations
- Proposed framework improves computational and memory efficiency compared to previous techniques relying on 3D CNNs
- Outperforms existing methods based on 3D CNNs in terms of false positive reduction in pulmonary nodule detection
- Utilizing ensemble of 2D CNNs with single views effectively addresses challenges associated with reducing false positives
- Research presents a novel approach to improve accuracy of pulmonary nodule detection using ensemble of 2D CNNs and automatic categorization of non-nodules based on autoencoder and k-means clustering
- Proposed framework shows promising results in reducing false positives and offers computational and memory efficiency advantages over existing methods.

## Summary

The study focuses on improving the accuracy of pulmonary nodule detection by reducing false positives. The first stage of nodule detection involves identifying suspicious nodules, but this often results in a high number of false positives. To address this issue, the researchers propose a novel framework that utilizes an ensemble of 2D convolutional neural networks (CNNs) with single views. One challenge in reducing false positives is the imbalance between the numbers of nodules and non-nodules, as well as the intra-class diversity of non-nodules. While 3D CNNs have shown promising performance, they suffer from high computational complexity, making it difficult to construct deep networks. To overcome these problems, the researchers introduce an ensemble of 2D CNNs using single-view 2D patches. In their approach, non-nodules are categorized automatically using an autoencoder and k-means clustering, eliminating the need for manual categorization by radiologists. The researchers train all 2D CNNs using the same nodule samples but with different types of non-nodules. This training scheme extends the learning capability and resolves difficulties in extracting representative features from non-nodules with large appearance variations. The proposed framework improves both computational and memory efficiency compared to previous techniques that rely on 3D CNNs. It outperforms existing methods based on 3D CNNs in terms of false positive reduction in pulmonary nodule detection. The findings suggest that utilizing an ensemble of 2D CNNs with single views can effectively address the challenges associated with reducing false positives in pulmonary nodule detection. Overall, this research presents a novel approach to improve the accuracy of pulmonary nodule detection by utilizing an ensemble of 2D CNNs and automatic categorization of non-nodules based on an autoencoder and k-means clustering. The proposed framework shows promising results in reducing false positives and offers computational and memory efficiency advantages over existing methods.