[Paper Review] A Pulmonary Nodule Detection Model Based on Progressive Resolution and Hierarchical Saliency

Junjie Zhang, Yong Xia+1 more|arXiv (Cornell University)|Jul 2, 2018

Lung Cancer Diagnosis and Treatment11 references3 citations

TL;DR

This paper proposes a two-stage deep learning framework for pulmonary nodule detection on chest CT scans, combining a 3D progressive resolution-based fully convolutional network (PRN) to detect candidates and a hierarchical saliency 3D CNN (HSN) to classify true nodules and estimate their diameters. The method achieved a state-of-the-art FROC score of 0.958 on the LUNA16 benchmark, outperforming all prior entries on the leaderboard.

ABSTRACT

Detection of pulmonary nodules on chest CT is an essential step in the early diagnosis of lung cancer, which is critical for best patient care. Although a number of computer-aided nodule detection methods have been published in the literature, these methods still have two major drawbacks: missing out true nodules during the detection of nodule candidates and less-accurate identification of nodules from non-nodule. In this paper, we propose an automated pulmonary nodule detection algorithm that jointly combines progressive resolution and hierarchical saliency. Specifically, we design a 3D progressive resolution-based densely dilated FCN, namely the progressive resolution network (PRN), to detect nodule candidates inside the lung, and construct a densely dilated 3D CNN with hierarchical saliency, namely the hierarchical saliency network (HSN), to simultaneously identify genuine nodules from those candidates and estimate the diameters of nodules. We evaluated our algorithm on the benchmark LUng Nodule Analysis 2016 (LUNA16) dataset and achieved a state-of-the-art detection score. Our results suggest that the proposed algorithm can effectively detect pulmonary nodules on chest CT and accurately estimate their diameters.

Motivation & Objective

• To address the limitations of existing CADe systems in missing true nodules and misclassifying non-nodules.
• To improve nodule candidate detection by leveraging multi-scale, progressive resolution feature learning.
• To enhance nodule verification and diameter estimation using hierarchical saliency features in a 3D CNN.
• To achieve state-of-the-art performance on the LUNA16 benchmark dataset.
• To enable accurate, real-time detection suitable for clinical deployment.

Proposed method

• The progressive resolution network (PRN) uses a 3D densely dilated FCN with multi-scale receptive fields to detect nodule candidates across varying resolutions.
• The hierarchical saliency network (HSN) employs a 3D CNN with hierarchical feature learning to classify candidates as true nodules and estimate their diameters.
• The HSN uses a 3D transition layer with dense connections and a 3D DDB block to capture multi-level contextual features.
• A regression head with mean absolute error loss estimates nodule diameter, while a classification head uses a linear output for nodule genuineness.
• Data augmentation via rotation and flipping increased positive sample diversity, with a 1:10 ratio of positive to negative samples during training.
• Testing used a sliding window of 32×32×32 voxels over the lung, with majority voting and averaging across augmented predictions for final detection.

Experimental results

Research questions

• RQ1 Can a progressive resolution strategy improve the detection of small and subtle pulmonary nodule candidates in 3D CT scans?
• RQ2 Can hierarchical saliency features enhance the discrimination between true nodules and false positives in a 3D CNN?
• RQ3 Does the joint use of candidate detection and refined classification improve overall nodule detection performance on benchmark datasets?
• RQ4 Can the proposed method achieve superior performance compared to existing state-of-the-art methods on the LUNA16 dataset?
• RQ5 Is the method computationally efficient enough for real-time clinical deployment?

Key findings

• The proposed method achieved a FROC score of 0.958 on the LUNA16 dataset, surpassing the previous best record of 0.951.
• The average absolute difference between detected and true nodule diameters was 1.34 mm, indicating high accuracy in size estimation.
• The method successfully detected small nodules with diameters below 5 mm, demonstrating sensitivity to early-stage lesions.
• The HSN model effectively reduced false positives by leveraging hierarchical saliency to distinguish subtle visual differences between nodules and non-nodule tissues.
• Training took approximately 18 hours for PRN and 23 hours for HSN, with inference taking only about 3 minutes per scan on standard hardware.
• The model outperformed all top five entries on the LUNA16 Challenge leaderboard, establishing a new state-of-the-art result.