[Paper Review] False Positive Reduction in Lung Computed Tomography Images using Convolutional Neural Networks

Görkem Polat, Uğur Halıcı+1 more|arXiv (Cornell University)|Nov 4, 2018

Lung Cancer Diagnosis and Treatment10 references4 citations

TL;DR

This paper proposes a 3D convolutional neural network (CNN)-based framework for reducing false positives in lung computed tomography (CT) screening, leveraging 3D data to preserve spatial information and improve nodule classification. The method achieves a mean sensitivity of 0.786 at seven predefined false positive rates on the LUNA16 challenge, demonstrating that even simple 3D CNN architectures outperform traditional methods in false positive reduction.

ABSTRACT

Recent studies have shown that lung cancer screening using annual low-dose computed tomography (CT) reduces lung cancer mortality by 20% compared to traditional chest radiography. Therefore, CT lung screening has started to be used widely all across the world. However, analyzing these images is a serious burden for radiologists. In this study, we propose a novel and simple framework that analyzes CT lung screenings using convolutional neural networks (CNNs) and reduces false positives. Our framework shows that even non-complex architectures are very powerful to classify 3D nodule data when compared to traditional methods. We also use different fusions in order to show their power and effect on the overall score. 3D CNNs are preferred over 2D CNNs because data are in 3D, and 2D convolutional operations may result in information loss. Mini-batch is used in order to overcome class-imbalance. Proposed framework has been validated according to the LUNA16 challenge evaluation and got score of 0.786, which is the average sensitivity values at seven predefined false positive (FP) points.

Motivation & Objective

• To address the high false positive rate in lung CT screening, which burdens radiologists and reduces diagnostic reliability.
• To develop a deep learning framework that effectively classifies 3D pulmonary nodules using 3D CNNs to minimize false positives.
• To evaluate the performance of different feature fusion strategies in improving classification accuracy.
• To validate the framework on the standardized LUNA16 benchmark dataset for fair comparison with existing methods.

Proposed method

• The framework employs 3D convolutional neural networks to process volumetric CT data, preserving spatial context lost in 2D approaches.
• It uses mini-batch training to mitigate class imbalance, where benign nodules significantly outnumber malignant ones.
• Multiple fusion strategies are applied to combine features from different network branches, enhancing discriminative power.
• The model is trained and evaluated on the LUNA16 dataset, using standard evaluation metrics including sensitivity at predefined false positive rates.
• The architecture is designed to be simple yet effective, focusing on the utility of 3D convolutions over 2D for volumetric data.
• Evaluation is performed using the official LUNA16 challenge protocol, measuring mean sensitivity across seven false positive points.

Experimental results

Research questions

• RQ1 Can a simple 3D CNN architecture achieve superior false positive reduction compared to traditional methods in lung CT nodule classification?
• RQ2 How does the use of 3D convolutions compare to 2D convolutions in preserving diagnostic information from volumetric CT scans?
• RQ3 What is the impact of different feature fusion strategies on the overall performance of the false positive reduction framework?
• RQ4 To what extent can mini-batch training effectively address class imbalance in lung nodule classification?
• RQ5 How does the proposed framework perform on the standardized LUNA16 benchmark compared to state-of-the-art methods?

Key findings

• The proposed 3D CNN framework achieved a mean sensitivity of 0.786 at seven predefined false positive points on the LUNA16 challenge, indicating strong performance in false positive reduction.
• The use of 3D convolutions preserved spatial information better than 2D operations, reducing information loss and improving classification accuracy.
• Even non-complex 3D CNN architectures outperformed traditional methods in distinguishing malignant from benign nodules.
• Feature fusion strategies contributed positively to the overall performance, demonstrating their value in enhancing model robustness.
• The mini-batch training approach effectively mitigated class imbalance, improving model generalization on rare malignant nodules.
• The framework demonstrated strong generalization on the LUNA16 benchmark, confirming its potential for real-world clinical deployment.