Source-linked AI summary
NoduleNet: Decoupled False Positive Reductionfor Pulmonary Nodule Detection and Segmentation
Hao Tang, Chupeng Zhang, Xiaohui Xie
TL;DR
Nodule detection, false positive reduction, and segmentation are often handled separately, despite the costs and lost feature sharing of separate training. NoduleNet combines them in an end-to-end 3D network; on LIDC it reports improved detection accuracy and state-of-the-art detection and segmentation scores.
Problem
Separately trained task-specific networks can be resource-intensive and may miss feature sharing across detection, false positive reduction, and segmentation.
Source: §1 · p. 2Method
NoduleNet jointly performs candidate screening, false positive reduction, and segmentation using an end-to-end 3D network with decoupled features and segmentation refinement.
Source: §1 · p. 2, Decoupled false positive reduction… · p. 4, Segmentation refinement (SR) · p. 4Results
Takeaways & Limitations
The authors state that the techniques introduced can be readily transferred to other models.
Source: §4 · p. 6
Abstract
from arXiv · showhide
Pulmonary nodule detection, false positive reduction and segmentation represent three of the most common tasks in the computeraided analysis of chest CT images. Methods have been proposed for eachtask with deep learning based methods heavily favored recently. However training deep learning models to solve each task separately may be sub-optimal - resource intensive and without the benefit of feature sharing. Here, we propose a new end-to-end 3D deep convolutional neural net (DCNN), called NoduleNet, to solve nodule detection, false positive reduction and nodule segmentation jointly in a multi-task fashion. To avoid friction between different tasks and encourage feature diversification, we incorporate two major design tricks: 1) decoupled feature maps for nodule detection and false positive reduction, and 2) a segmentation refinement subnet for increasing the precision of nodule segmentation. Extensive experiments on the large-scale LIDC dataset demonstrate that the multi-task training is highly beneficial, improving the nodule detection accuracy by 10.27%, compared to the baseline model trained to only solve the nodule detection task. We also carry out systematic ablation studies to highlight contributions from each of the added components. Code is available at https://github.com/uci-cbcl/NoduleNet.
1 Introduction
NoduleNet jointly addresses nodule detection, false positive reduction, and segmentation, tasks that separately trained systems may handle inefficiently and without feature sharing. The paper reports improved detection accuracy and segmentation performance, and introduces decoupled features and a segmentation refinement subnet.
- Accurate segmentation can support nodule shape, size, and growth-rate assessment, including comparison with follow-up scans.A follow-up scan in 3–12 months is usually required to assess growth rate.Source: §1 · pp. 1–2
- Separately trained systems are time- and resource-intensive and may miss useful feature sharing across tasks.Existing CAD systems often use independent subsystems optimized separately.Source: §1 · p. 2
- Decoupling localization and classification features addresses mismatched task goals while retaining some shared feature extraction.The proposed false positive reduction pools early-scale features to learn task-independent and task-dependent representations.Source: §1 · p. 2
- NoduleNet jointly performs candidate screening, false positive reduction, and segmentation refinement using a shared backbone and end-to-end training.The overview depicts these components as three sequential stages.Source: §1 · p. 2, Figure 1 · p. 3
- Nodule detection accuracy improved 10.27% over the detection-only baseline, with nodule segmentation reaching 83.10% DSC.Source: §1 · p. 3
- Systematic ablations examine decoupled feature maps, the segmentation refinement subnet, and multi-task training.Source: §1 · p. 3
2 NoduleNet
NoduleNet screens candidate nodules, reduces false positives, and refines segmentation through three connected network components. The design uses distinct feature processing for false positive reduction and full-resolution refinement for segmentation.
Nodule candidate screening (NCS)
- NCS applies a 3 × 3 × 3 convolution followed by parallel 1 × 1 × 1 layers for anchor classification and regression.The five anchors are cubes of sizes 5, 10, 20, 30, and 50.Source: Nodule candidate screening (NCS) · pp. 3–4
Decoupled false positive reduction (DFPR)
- DFPR pools early feature-map regions with a small receptive field, separating its representations from screening while sharing some feature-extraction blocks.It uses 3D ROI pooling.Source: Decoupled false positive reduction… · p. 4
Segmentation refinement (SR)
- At input scale, SR upsamples cropped high-level features and concatenates low-level features, unlike 's downsampled-mask resize approach.The latter may lose precision through bounding-box regression errors and loss of fine-grained local features.Source: Segmentation refinement (SR) · p. 4
- SR upsamples only nodule regions to original image scale, saving GPU memory and making whole-volume input feasible during training and testing.Source: Segmentation refinement (SR) · p. 4
3 Results
NoduleNet was evaluated on LIDC data using six-fold cross-validation, with detection and segmentation assessed using task-specific metrics. Its final detection comparison and component analyses report improvements over the baseline and coupled variants, while segmentation outperformed a prior method.
Data and experiment configurations
- Evaluation used 586 CT scans with 1,131 nodules selected under LUNA16 criteria and six-fold cross-validation.Detection was evaluated with FROC/CPM; segmentation was evaluated with IoU and DSC.Source: Data and experiment configurations · pp. 4–5
Nodule detection performance
- Adding segmentation refinement improved sensitivity at 8 false positives per patient by 1.0% to 1.5%, and fusing NCS and FPR predictions improved performance by 0.7% to 1.0%.Adding false-positive reduction and segmentation refinement also improved the baseline NCS detector.Source: Nodule detection performance · p. 5
- Decoupled FPR improved average sensitivity by around 3% to 4% over coupled FPR; rotation augmentation further improved the FPR branch by around 2.5% while NCS remained almost unchanged.The passage relates these findings to classification learning invariant features and localization learning co-variant features.Source: Nodule detection performance · p. 5
- 10.27%: NoduleNet outperformed a strong baseline single-stage detector on nodule detection.The detection comparison is reported alongside Table 1's CPM sensitivities across false-positive thresholds.Source: Nodule detection performance · p. 6, Table 1 · p. 6
Nodule segmentation performance
- On the LIDC dataset, NoduleNet outperformed a previous state-of-the-art deep-learning-based method by 0.95% on DSC without needing to train a separate, dedicated 3D DCNN for nodule segmentation.Source: Nodule segmentation performance · p. 6
4 Conclusion
NoduleNet jointly addresses pulmonary nodule detection, false-positive reduction, and segmentation. On LIDC, it achieved reported state-of-the-art detection and segmentation scores.
4 Conclusion
- On LIDC, NoduleNet achieved a CPM score of 87.27% for nodule detection and a DSC score of 83.10% for nodule segmentation, reported as current state-of-the-art results on the dataset.The paper also reports systematic analysis of the architecture's components.Source: §4 · p. 6
- Table 2 reports IoU and DSC segmentation performance across methods, with consensus indicating the minimum number of annotating experts.A method includes nodules annotated by at least the specified number of experts.Source: Table 2 · p. 7
- Figure 2 presents examples of nodule segmentation generated by NoduleNet.Source: Figure 2 · p. 7