Source-linked AI summary
Automatic Pulmonary Nodule Detection in CT Scans Using Convolutional Neural Networks Based on Maximum Intensity Projection
Sunyi Zheng, Jiapan Guo, Xiaonan Cui, Raymond N. J. Veldhuis, Matthijs Oudkerk, Peter M. A. van Ooijen
TL;DR
Few CAD approaches account for radiologists’ routine use of MIP images during nodule screening. This paper combines multi-view MIP inputs with CNNs, reporting that merging four groups’ results raises sensitivity from 82.80% to 95.36%.
Problem
Few existing CAD approaches account for radiologists’ routine MIP workflow for locating candidates before examining specific CT slices.
Source: §I · p. 2Method
The system combines multiple MIP views with CNNs, adds spatial information to 2-D images, and fuses predictions from multiple networks.
Source: §I · p. 2Results
Sensitivity rose from 82.80% to 95.36% when results from four groups were merged; the system also showed strong detection with a low false-positive rate against LIDC/IDRI methods.
Source: §V · p. 7Takeaways & Limitations
The authors conclude that MIP images can benefit CNN-based detection, especially for small lesions, and that the system can reduce the scans radiologists need to evaluate.
Source: §VI · p. 8
Abstract
from arXiv · showhide
Accurate pulmonary nodule detection is a crucial step in lung cancer screening. Computer-aided detection (CAD) systems are not routinely used by radiologists for pulmonary nodule detection in clinical practice despite their potential benefits. Maximum intensity projection (MIP) images improve the detection of pulmonary nodules in radiological evaluation with computed tomography (CT) scans. Inspired by the clinical methodology of radiologists, we aim to explore the feasibility of applying MIP images to improve the effectiveness of automatic lung nodule detection using convolutional neural networks (CNNs). We propose a CNN-based approach that takes MIP images of different slab thicknesses (5 mm, 10 mm, 15 mm) and 1 mm axial section slices as input. Such an approach augments the two-dimensional (2-D) CT slice images with more representative spatial information that helps discriminate nodules from vessels through their morphologies. Our proposed method achieves sensitivity of 92.67% with 1 false positive per scan and sensitivity of 94.19% with 2 false positives per scan for lung nodule detection on 888 scans in the LIDC-IDRI dataset. The use of thick MIP images helps the detection of small pulmonary nodules (3 mm-10 mm) and results in fewer false positives. Experimental results show that utilizing MIP images can increase the sensitivity and lower the number of false positives, which demonstrates the effectiveness and significance of the proposed MIP-based CNNs framework for automatic pulmonary nodule detection in CT scans. The proposed method also shows the potential that CNNs could gain benefits for nodule detection by combining the clinical procedure.
I. INTRODUCTION
Pulmonary CAD faces a difficult detection task because nodules vary in size and morphology, and few approaches incorporate radiologists’ routine use of MIP images. The paper proposes a multi-view MIP-based CNN framework that follows this clinical workflow.
I. INTRODUCTION
- Early lung-cancer detection is described as crucial because it could improve survival chances; the cited five-year survival rate is 16%.The introduction characterizes lung cancer as a severe, high-incidence disease.Source: §I · p. 1
- Pulmonary CAD must balance finding as many nodule candidates as possible against eliminating false positives without missing nodules.Nodule complexity, including differences in size, shape, margins, and calcification, makes detection difficult.Source: §I · p. 1
- Few prior CAD approaches account for radiologists’ routine first look at MIP images to locate candidates before examining specific slices.MIP projects maximum-intensity voxels from 3-D data onto a plane and enhances nodule visualization relative to bronchi and vasculature.Source: §I · p. 2
- The framework combines MIP views to add spatial information for distinguishing nodules from vessels, then uses multiple CNNs to reduce false positives and fuse predictions.The authors also present it as more interpretable to radiologists and as a feasibility demonstration of combining clinical screening practice with CNNs.Source: §I · p. 2
II. MATERIALS
The study uses the LIDC/IDRI scans retained for LUNA16 and selects consensus nodules at least 3 mm in diameter for evaluation.
A. Database
- The LIDC/IDRI repository contains 1,018 thoracic CT scans, and LUNA16 retained 888 after excluding scans with slice thickness above 2.5 mm.The dataset slice thicknesses ranged from 0.6 mm to 5.0 mm; the exclusion followed guidance against thicker slices for clinical nodule screening.Source: A. Database · p. 2
- Four experienced radiologists annotated nodules in a two-stage reading process, and nodules accepted by at least three readers formed the reference standard.The resulting experiments included 1,186 valid nodules at least 3 mm in diameter.Source: A. Database · p. 2
- Figure 2 presents the lung-parenchyma segmentation sequence from a raw thoracic CT image through successive masks to the segmented image.Source: Figure 2 · p. 3
III. METHODS
The detection system first generates nodule candidates from segmented lung images, then classifies candidate-centered cubic patches to reduce false positives.
III. METHODS
- The system has two stages: four 2-D CNNs detect candidates from MIP images, and an ensemble of 3-D CNNs classifies candidate-centered cubic patches.Detections from the four candidate-detection networks are merged before the second-stage classification.Source: §III · p. 3
A. Lung Parenchyma Segmentation
- Segmentation removes irrelevant chest regions by thresholding, removing isolated border-connected pixels, and selecting the two largest internal chest components.Images are windowed from -1000 HU to 400 HU and normalized to 0–1; closing and dilation preserve boundary information, while binary hole filling removes small noisy regions.Source: A. Lung Parenchyma Segmentation · p. 3
B. Maximum Intensity Projection
The method supplies 1 mm axial slices and MIP images at three slab thicknesses to CNN streams, combining their contextual information for candidate detection and later classification.
B. Maximum Intensity Projection
- The input set comprises 1 mm axial slices and MIP images with slab thicknesses of 5 mm, 10 mm, and 15 mm.MIP superimposes maximum grey values at each coordinate across consecutive slices; the chosen slab thicknesses are described as clinically beneficial for nodule detection.Source: B. Maximum Intensity Projection · p. 3
C. MIP-based Convolutional Networks and Model Fusion
- The four-stream candidate-detection system is followed by false-positive reduction that assigns candidate locations a probability of being a nodule.Figure 4 depicts the input images, candidate fusion, probability output, and network architectures.Source: Figure 4 · p. 4
- Four 2-D CNN streams use images at different slab thicknesses to capture distinct context and distinguish nodules from vessels through morphology.The streams’ predictions are fused into nodule candidates.Source: C. MIP-based Convolutional Networks… · p. 4, Figure 4 · p. 4
- The candidate-localization network is U-net-based, with encoding and decoding parts and a final 1 × 1 convolution followed by sigmoid activation.The described network has 18 convolutional layers.Source: C. MIP-based Convolutional Networks… · pp. 4–5
- The candidate-detection-stage evaluation covers 888 scans and 1,186 nodules, including 905 nodules sized 3–10 mm.The caption also lists 231 nodules sized 10–20 mm and 50 sized at least 20 mm.Source: Table I · p. 5
D. False Positive Reduction
The false-positive reduction stage uses 3-D CNNs to classify merged nodule candidates, with two architectures that take patches at different sizes. Their designs reflect the candidate-size analysis and use VGG-net-based layers.
D. False Positive Reduction
- The 3-D CNNs classify merged candidates as nodule or non-nodule, using 16- and 32-pixel cubes based on candidate sizes.The networks are based on VGG-net and discriminate nodules from irrelevant findings.Source: D. False Positive Reduction · p. 5
- Archi-3 uses 16 × 16 × 16-pixel patches and 12 layers, with three repeated convolution, max-pooling, and batch-normalization applications.Each application has two 3 × 3 × 3 convolutional layers with 16 kernels, and kernel counts double across applications.Source: D. False Positive Reduction · p. 5
- Archi-2 follows a similar structure to Archi-3 but adds three convolutional layers, a max-pooling layer, and batch normalization for its larger patches.Its maximum number of kernels is 128.Source: D. False Positive Reduction · p. 5
E. Training Process
Training used 10-fold cross-validation with separate settings for candidate detection and false-positive reduction. The process also used data augmentation, Adam optimization, and early stopping for candidate detection.
E. Training Process
- Candidate detection used 10-fold cross-validation with 63% of data for training, 27% for validation, and 10% for testing, plus early stopping with patience 10.The batch size was 5, and learning rates ranged from 10^-3 to 10^-7 with a reduction factor of 0.01.Source: E. Training Process · p. 5
- Candidate-detection training used Dice score coefficient loss, He initialization, Adam optimization, and translated, rotated, and flipped nodule-containing patches for augmentation.Patches were translated up to 30 voxels along each axis and rotated by randomly selected multiples of 90 degrees.Source: E. Training Process · p. 5
- False-positive reduction used the same cross-validation proportions, trained on imbalanced candidates, and fused size-specific Archi-2 and Archi-3 probabilities by weighted linear combination.The first stage produced 15 times as many false-positive nodules as true nodules; the weights were determined by grid search.Source: E. Training Process · p. 6
A. Candidates Detection
At the candidate-detection stage, combining axial slices with MIP streams produced higher sensitivity than any individual stream. Thicker MIPs particularly improved detection of small nodules and reduced false positives relative to axial slices.
A. Candidates Detection
- 95.36% sensitivity with 19.13 false positives per scan was achieved by the four-stream merged architecture.The four streams used 1 mm axial slices and 5 mm, 10 mm, and 15 mm MIP images.Source: A. Candidates Detection · p. 6, Table II · p. 6
- 10 mm MIP detected the most small nodules with a relatively low false-positive rate, while thicker MIP streams improved sensitivity for 3–10 mm nodules and had fewer false positives than axial slices.All four streams detected similar numbers of nodules larger than 10 mm, and none missed nodules larger than 20 mm.Source: A. Candidates Detection · p. 6
B. False Positive Reduction
The FROC evaluation reports performance across low false-positive rates, with fusion maintaining sensitivity as the rate falls from 16 to 2 false positives per scan. All single architectures exceeded 90.75% sensitivity at 1 false positive per scan.
B. False Positive Reduction
- 89.89% sensitivity at 0.25 false positives per scan and 94.81% at 4 false positives per scan were obtained by fusion.As the false-positive rate decreased from 16 to 2 per scan, sensitivity remained steady.Source: B. False Positive Reduction · p. 6
- All single architectures achieved over 90.75% sensitivity at 1 false positive per scan.The FROC evaluation covered 0.125, 0.25, 0.5, 1, 2, 4, 8, and 16 false positives per scan.Source: B. False Positive Reduction · p. 6
V. DISCUSSION
The discussion presents multi-slab MIP as a way to bring radiologists’ screening workflow into CNN-based nodule detection. It reports comparative performance and describes remaining detection and evaluation gaps.
V. DISCUSSION
- On 888 LIDC/IDRI scans, the method had results comparable to one system using two CNN types, required less processing time, and Zhang et al. had more candidate-stage false positives.The system was designed for all nodules, whereas Setio et al. included systems targeting special nodule types and combined multiple detection and false-positive-reduction systems.Source: Table III · p. 7, §V · p. 7
- Merging results from four groups improved performance from a lowest sensitivity of 82.80% to 95.36%.The system had the fewest candidates per scan with sensitivity comparable to other systems.Source: §V · p. 7
- Thicker MIP images were associated with fewer false positives because they help distinguish nodules from vessels.The discussion also says MIP images provide structural spatial information to 2-D images, while 3-D candidate-classification networks are ensembled to remove false positives.Source: §V · p. 7
- The 2-D localization approach uses less computing power and storage than 3-D CNNs, while MIP adds spatial context for distinguishing nodules from vessels.More accurate localization lets the later classification stage analyze smaller cubic patches with affordable computation.Source: §V · p. 7
- The authors identify room to improve detection and false-positive rates and call for evaluation on more screening-trial data and low-dose CT scans.They also note that detection of small and ground nodules needs improvement.Source: §V · p. 7
VI. CONCLUSION
The conclusion presents a lung-nodule CAD system using MIP views with varied slab thicknesses. It reports benefits for false-positive reduction and small-lesion detection, while identifying further clinical-data validation as future work.
VI. CONCLUSION
- The system applies multiple MIP views with varied slab thicknesses to lung nodule detection.The conclusion describes integrating multi-slab MIP images for CT-based detection.Source: §VI · p. 7
- The authors report accurate, robust performance for false-positive reduction and benefits for detecting small pulmonary lesions.They describe combining networks at various scales.Source: §VI · p. 8
- The system is proposed as a way to reduce the scans radiologists need to evaluate, and the authors say it requires further validation on other clinical data.Promotion in lung cancer screening is framed as a future possibility.Source: §VI · p. 8