Source-linked AI summary
Accurate Pulmonary Nodule Detection in Computed Tomography Images Using Deep Convolutional Neural Networks
Jia Ding, Aoxue Li, Zhiqiang Hu, Liwei Wang
TL;DR
Pulmonary cancer screening requires accurate pulmonary nodule detection, while traditional CAD approaches struggle with variable nodule appearance and limited 3D context. The paper combines deconvolutional Faster R-CNN candidate detection with 3D DCNN false-positive reduction, achieving first place on LUNA16 with a 0.891 average FROC-score.
Problem
Traditional CAD systems rely on low-level descriptors that fail to capture variable nodule shape, size, and texture, while direct nodule detection rarely incorporates CT's 3D context.
Method
The CAD system uses deconvolutional Faster R-CNN for candidate detection on axial slices followed by a 3D DCNN for false positive reduction.
Results
0.891 average FROC-score placed the proposed system first in the LUNA16 Nodule Detection Track.
Takeaways & Limitations
The proposed DCNN-based CAD system demonstrates high pulmonary nodule detection performance on the LUNA16 challenge.
Takeaways & Limitations
Traditional approaches are constrained by variability in nodule shape, size, and texture and by limited use of 3D context in direct detection.
Abstract
from arXiv · showhide
Early detection of pulmonary cancer is the most promising way to enhance a patient's chance for survival. Accurate pulmonary nodule detection in computed tomography (CT) images is a crucial step in diagnosing pulmonary cancer. In this paper, inspired by the successful use of deep convolutional neural networks (DCNNs) in natural image recognition, we propose a novel pulmonary nodule detection approach based on DCNNs. We first introduce a deconvolutional structure to Faster Region-based Convolutional Neural Network (Faster R-CNN) for candidate detection on axial slices. Then, a three-dimensional DCNN is presented for the subsequent false positive reduction. Experimental results of the LUng Nodule Analysis 2016 (LUNA16) Challenge demonstrate the superior detection performance of the proposed approach on nodule detection(average FROC-score of 0.891, ranking the 1st place over all submitted results).
1 Introduction
Pulmonary cancer is a major global health burden, and expanding CT screening has increased the volume of images radiologists must analyze. The paper proposes a two-stage DCNN-based CAD system to improve pulmonary nodule detection.
- Pulmonary cancer causes 1.3 million deaths annually and is a leading cause of cancer death worldwide.
- CT mass screening enables earlier tumor detection but creates a rapidly growing image-analysis workload for radiologists.
- Traditional CAD systems use simple shape assumptions and hand-crafted low-level descriptors that fail to capture highly variable nodule features.
- Because CT images are inherently three-dimensional, limited use of 3D context remains an issue in direct nodule detection.
- The proposed CAD system detects candidates with deconvolutional Faster R-CNN and reduces false positives with a 3D DCNN.
- 0.891 average FROC-score ranked the system first in the LUNA16 Nodule Detection Track.
2 The proposed CAD system
The proposed CAD system uses a two-stage DCNN pipeline: an improved Faster R-CNN detects candidates on axial slices, followed by a 3D DCNN that reduces false positives. The design adds fine-grained features for small nodules and incorporates 3D candidate context.
- Proposed CAD system: The CAD system first detects nodule candidates with Faster R-CNN and then reduces false positives using a three-dimensional DCNN.These are the two main stages of the proposed system.
- Candidate detection: Axial-slice inputs concatenate two neighboring slices and are rescaled to 600×600×3 pixels for candidate detection.Using slices instead of full 3D volumes reduces computation cost.
- Candidate detection: A deconvolutional layer recovers finer-grained features before the region proposal network generates regions of interest and predicts their objectness scores.The added layer uses kernel size 4, stride 4, padding 2, and 512 kernels.
- Candidate detection: Six anchors sized 4×4, 6×6, 10×10, 16×16, 22×22, and 32×32 are designed to fit pulmonary nodules of different sizes.Multiple region proposals are predicted at each sliding-window location.
- Candidate detection: ROI pooling maps each proposal to a fixed 7×7 feature map before fully connected layers classify it as a nodule or not.Pooling divides each proposal into a grid and max-pools each sub-window independently across feature-map channels.
- False positive reduction: The 3D DCNN uses six 3D convolutional layers, three 3D max-pooling layers, three fully connected layers, and a final two-way softmax.For each candidate, the network processes a normalized 40×40×24 CT patch, with cropping and flipping used for augmentation.
3 Experimental Results and Discussions
The CAD system was evaluated on the 888-scan LUNA16 dataset using FROC analysis, with candidate detection and false-positive reduction assessed against baselines and submitted systems. It achieved strong candidate sensitivity and the best reported overall performance, while competing systems dropped at lower false-positive rates.
- Evaluation setting: The evaluation used 888 CT scans from LIDC-IDRI after excluding scans with excessive slice thickness, inconsistent spacing, or missing slices.
- Evaluation setting: FROC analysis plots sensitivity against average false positives per scan, with average FROC-score averaging sensitivities at seven specified false-positive rates.
- Candidate detection: 94.6% sensitivity with 15.0 candidates per scan was achieved by the proposed CAD system, the highest sensitivity with the fewest candidates among compared systems.
- Candidate detection: The deconvolutional Faster R-CNN outperformed its version without deconvolution, while six anchors were reported as more suitable than four anchors for candidate detection.
- False-positive reduction: The proposed 3D DCNN significantly outperformed the 2D NIN baseline in false-positive reduction and attained the best performance among submitted CAD systems.
- False-positive reduction: Other systems showed a significant performance drop below 2 false positives per scan, whereas the authors state their system fits clinical operation between 1 and 4 false positives per scan.
4 Conclusion
The study proposes a deep-convolutional CAD system combining deconvolution-enhanced Faster R-CNN candidate detection with 3D DCNN false-positive reduction. On the LUNA16 Nodule Detection Challenge, it ranked first with an average FROC-score of 0.891.
- The CAD system combines deconvolution-enhanced Faster R-CNN for axial-slice candidate detection with a 3D DCNN for false-positive reduction.
- 0.891 average FROC-score placed the proposed system first in the LUNA16 Nodule Detection Track.