[Paper Review] RUN:Residual U-Net for Computer-Aided Detection of Pulmonary Nodules without Candidate Selection

Tian Lan, Yuanyuan Li+3 more|arXiv (Cornell University)|May 30, 2018

Lung Cancer Diagnosis and TreatmentMedicine25 references20 citations

TL;DR

This paper proposes RUN, a residual U-Net architecture that enables end-to-end, single-step detection of pulmonary nodules in low-dose CT scans without requiring candidate selection. By integrating residual learning into the U-Net encoder-decoder structure, RUN achieves a 90.90% sensitivity at 2 false positives per scan on the LUNA16 dataset, outperforming state-of-the-art methods.

ABSTRACT

The early detection and early diagnosis of lung cancer are crucial to improve the survival rate of lung cancer patients. Pulmonary nodules detection results have a significant impact on the later diagnosis. In this work, we propose a new network named RUN to complete nodule detection in a single step by bypassing the candidate selection. The system introduces the shortcut of the residual network to improve the traditional U-Net, thereby solving the disadvantage of poor results due to its lack of depth. Furthermore, we compare the experimental results with the traditional U-Net. We validate our method in LUng Nodule Analysis 2016 (LUNA16) Nodule Detection Challenge. We acquire a sensitivity of 90.90% at 2 false positives per scan and therefore achieve better performance than the current state-of-the-art approaches.

Motivation & Objective

• To develop a deep learning framework that eliminates the need for candidate selection in pulmonary nodule detection.
• To improve detection accuracy for small and variable-sized nodules by enhancing U-Net's depth and training stability.
• To achieve end-to-end nodule detection directly from volumetric CT scans using a single, unified network.
• To validate the method on the LUNA16 benchmark and demonstrate clinical relevance through high sensitivity at low false positive rates.

Proposed method

• Integrates residual learning blocks into the U-Net encoder-decoder architecture to enable deeper network training and mitigate vanishing gradient problems.
• Uses skip connections (shortcuts) from residual networks to preserve feature maps and improve gradient flow during backpropagation.
• Processes entire CT slices as input instead of cropping patches, preserving spatial context and reducing false positives from spatial information loss.
• Employs a multi-scale training strategy using single- and multi-slice inputs to evaluate the impact of spatial context on detection performance.
• Trains the network end-to-end using cross-entropy loss and Dice coefficient for segmentation, optimizing for both localization and classification.
• Applies preprocessing steps including noise reduction, lung segmentation via threshold and morphological operations, and image normalization.

Experimental results

Research questions

• RQ1 Can a residual U-Net architecture outperform standard U-Net in detecting small and variable-sized pulmonary nodules?
• RQ2 Does eliminating the candidate selection stage improve detection performance and simplify the CAD pipeline?
• RQ3 How does the use of full-slice input versus patch-based input affect detection sensitivity and false positive rates?
• RQ4 Can the integration of residual learning into U-Net significantly improve segmentation accuracy and generalization on challenging nodules?

Key findings

• RUN achieved a sensitivity of 90.90% at 2 false positives per scan on the LUNA16 dataset, surpassing the current state-of-the-art methods.
• The network demonstrated a 6% improvement in Dice coefficient (71.9%) compared to the standard U-Net (66%), indicating superior segmentation performance.
• RUN outperformed both the standard U-Net and the dual-path residual U-Net in segmentation accuracy, especially for small nodules.
• The method achieved 79.05% sensitivity at 0.11 false positives per scan, indicating strong performance even at low false positive rates.
• Multi-slice training improved spatial context awareness but risked false learning for very small nodules, suggesting a need for size-adaptive slice selection.