Source-linked AI summary
Automated Pulmonary Nodule Detection via 3D ConvNets with Online Sample Filtering and Hybrid-Loss Residual Learning
Qi Dou, Hao Chen, Yueming Jin, Huangjing Lin, Jing Qin, Pheng-Ann Heng
TL;DR
Automated pulmonary nodule detection from low-dose CT is difficult because medical datasets contain severe hard/easy sample imbalance and localized annotations are underused. The paper proposes a two-stage 3D ConvNet framework with online sample filtering for candidate screening and a hybrid-loss residual network for false positive reduction. Experiments on LUNA16 report improved detection performance, while ablations show benefits from online filtering and location-and-size supervision.
Problem
Pulmonary nodule detection from low-dose CT is challenging because medical datasets contain severe hard/easy sample imbalance, while early detection is important for lung cancer diagnosis and treatment.
Method
A two-stage 3D ConvNet framework uses an online sample filtering 3D FCN for candidate screening and a hybrid-loss residual network using localization information for false positive reduction.
Results
The method outperformed state-of-the-art approaches on LUNA16, while ablations improved screening sensitivity from 94.3% to 97.1% and reduced FPs/scan from 286.2 to 219.1 with online sample filtering.
Takeaways & Limitations
Online filtering and localized location-and-size supervision improve the proposed ConvNet framework, which the authors describe as extensible to other medical image computing applications.
Abstract
from arXiv · showhide
In this paper, we propose a novel framework with 3D convolutional networks (ConvNets) for automated detection of pulmonary nodules from low-dose CT scans, which is a challenging yet crucial task for lung cancer early diagnosis and treatment. Different from previous standard ConvNets, we try to tackle the severe hard/easy sample imbalance problem in medical datasets and explore the benefits of localized annotations to regularize the learning, and hence boost the performance of ConvNets to achieve more accurate detections. Our proposed framework consists of two stages: 1) candidate screening, and 2) false positive reduction. In the first stage, we establish a 3D fully convolutional network, effectively trained with an online sample filtering scheme, to sensitively and rapidly screen the nodule candidates. In the second stage, we design a hybrid-loss residual network which harnesses the location and size information as important cues to guide the nodule recognition procedure. Experimental results on the public large-scale LUNA16 dataset demonstrate superior performance of our proposed method compared with state-of-the-art approaches for the pulmonary nodule detection task.
1 Introduction
Automated pulmonary nodule detection from low-dose CT is important but difficult because of large variation in nodule characteristics and the scale of screening data. The paper proposes a two-stage 3D ConvNet framework for candidate screening and false positive reduction.
- Motivation: Low-dose CT screening generates too much data for manual analysis to process cost-effectively.The paper motivates automated detection for high-risk population screening.
- Challenges: Pulmonary nodules are challenging to detect automatically because their diameters range from 3–30 mm and their shape, density, and anatomical context vary.These variations complicate reliable detection.
- Existing systems: Existing systems typically screen candidates sensitively in a first stage, then remove false positives in a second stage.The two-stage structure separates broad candidate retrieval from final discrimination.
- Existing systems: Earlier approaches commonly used hand-crafted image operations and low-level descriptors based on intensity, size, shape, texture, and context.These features were used across candidate screening and false positive reduction.
- Proposed framework: The proposed framework uses 3D ConvNets in both stages, with online sample filtering for screening and a hybrid-loss residual network for recognition.The design also uses location and size information to guide the second stage.
2 Method
The method uses a 3D fully convolutional network with online hard-sample filtering for rapid candidate screening, followed by a residual network that jointly learns classification and localization. Its hybrid loss uses localized annotations to improve recognition and estimate nodule size.
- Two-stage framework: The framework first applies a 3D FCN to screen candidates from volumetric CT scans, then uses a hybrid-loss 3D residual network to distinguish true nodules.The first stage prioritizes sensitivity and efficiency, while the second stage performs false positive reduction.
- Candidate screening: The 3D FCN is trained on small nodule and non-nodule patches but produces a full 3D score volume when applied to an entire CT scan.Candidate positions are retrieved from the score volume according to suspicious probabilities.
- Candidate screening: Online sample filtering addresses hard/easy imbalance by selecting high-loss samples during SGD while retaining some low-loss samples.The procedure dynamically increases the proportion of informative hard samples without interrupting learning or adding testing computations.
- Candidate screening: Candidate proposals are obtained by applying 3D nonmaximum suppression to the score volume and mapping the retained positions back to the original image space.The mapping accounts for the reduced score-volume dimensions induced by the network architecture.
- False positive reduction: The second-stage residual network jointly optimizes nodule classification and localization, sharing parameters in early layers.Localization uses annotations describing where a nodule is and how large it is.
- False positive reduction: The hybrid loss combines classification loss, localization loss, and weight decay, using a robust L1 function for localization and applying it only to positive samples.The localization target encodes scale-invariant translation and log-space size shifts relative to the cropped patch.
3 Experimental Results
On LUNA16, the framework was evaluated through cross-validation, component ablations, and comparisons across network configurations. Results show strong overall detection performance and gains from online sample filtering and hybrid-loss residual learning.
- Evaluation protocol: The LUNA16 evaluation used ten-fold cross-validation with sensitivity, FPs/scan, and CPM as detection metrics.A detection counted as true positive when it fell within a nodule centroid’s radius; CPM averaged sensitivity at seven predefined false-positive rates.
- Comparison with other methods: The method achieved a CPM of 0.839 and the best results at five of seven predefined FPs/scan rates.The comparison was conducted against other methods on the LUNA16 challenge dataset.
- Ablation studies: Online sample filtering increased screening sensitivity from 94.3% to 97.1% while reducing FPs/scan from 286.2 to 219.1.The ablation compared 3D FCN training with and without the online sample filtering scheme.
- Ablation studies: At 1.0 FPs/scan, sensitivity rose from 84.8% with DeepNet to 86.7% with ResNet and 90.5% with ResNet+HL.The hybrid-loss residual network additionally used location and size information during training.
- Ablation studies: ResNet+HL continually achieved the best performance among the three configurations across the FROC comparison.The figure also shows detection examples with prediction probabilities and regressed diameters.
4 Conclusion
The paper presents a two-stage 3D ConvNet framework for pulmonary nodule detection in low-dose CT. Its online filtering and hybrid-loss residual learning address sample imbalance and use localized information, with experiments validating the framework’s efficacy.
- Conclusion: The framework uses a 3D FCN with online sample filtering for candidate screening and a hybrid-loss residual ConvNet for false positive reduction.The second stage harnesses localized annotations to improve nodule classification accuracy.
- Conclusion: The method addresses hard/easy sample imbalance and incorporates localized annotations as learning guidance for pulmonary nodule detection.The conclusion identifies both issues as common in medical image computing with ConvNets.
- Conclusion: Extensive experiments validate the efficacy of the proposed detection framework.The reported conclusion is based on the framework’s experimental evaluation.