Source-linked AI summary
Discriminative Localization in CNNs for Weakly-Supervised Segmentation of Pulmonary Nodules
Xinyang Feng, Jie Yang, Andrew F. Laine, Elsa D. Angelini
TL;DR
Pulmonary nodule segmentation normally requires costly voxel-level annotations, motivating a weakly supervised alternative using image-level labels. The paper adapts a classification CNN to localize nodules at multiple scales and screens candidates with residual activation maps. On LIDC-IDRI, the resulting framework achieves competitive performance against a fully supervised CNN-based segmentation method while reducing annotation requirements.
Problem
Existing supervised nodule segmentation requires labor-intensive voxel-level annotations, whereas image-level labels can be obtained more efficiently.
Method
The method adapts a classification CNN to generate multi-scale nodule activation maps, coarse candidate segmentations, and residual-map-based fine segmentation.
Results
The weakly supervised framework achieves competitive performance compared with a CNN-based fully supervised segmentation method on the public LIDC-IDRI dataset.
Takeaways & Limitations
The framework reduces required labeling by approximately 100 times relative to voxel-based labels while retaining favorable detection and high average segmentation accuracy on detected nodules.
Takeaways & Limitations
The method currently targets one nodule per NAM, and future work proposes incorporating 3D contextual information and more annotated data for a dedicated network.
Abstract
from arXiv · showhide
Automated detection and segmentation of pulmonary nodules on lung computed tomography (CT) scans can facilitate early lung cancer diagnosis. Existing supervised approaches for automated nodule segmentation on CT scans require voxel-based annotations for training, which are labor- and time-consuming to obtain. In this work, we propose a weakly-supervised method that generates accurate voxel-level nodule segmentation trained with image-level labels only. By adapting a convolutional neural network (CNN) trained for image classification, our proposed method learns discriminative regions from the activation maps of convolution units at different scales, and identifies the true nodule location with a novel candidate-screening framework. Experimental results on the public LIDC-IDRI dataset demonstrate that, our weakly-supervised nodule segmentation framework achieves competitive performance compared to a fully-supervised CNN-based segmentation method.
1 Introduction
Pulmonary nodule segmentation could support earlier lung cancer diagnosis, but existing supervised methods require costly voxel-level annotations. This work adapts CNN classification to produce fully automated weakly supervised segmentation from binary slice-level labels and reports competitive performance against fully supervised segmentation.
- Motivation: Pulmonary nodules are CT-visible abnormalities regarded as crucial indicators of primary lung cancers.Their detection and segmentation can facilitate early diagnosis and timely surgical intervention.
- Research gap: Existing supervised nodule-segmentation methods require labor-intensive and time-consuming voxel-level annotations.Image-level labels indicating nodule presence can be obtained more efficiently.
- Research gap: Prior weakly labeled methods still require user inputs such as exact nodule location and estimated nodule size.
- Background: CNN classification models can identify discriminative regions through global-average-pooled activation maps, but medical nodules vary in size and may be subtle.This limitation motivates adapting localization methods for weakly supervised disease detection.
- Contribution: The proposed framework uses binary slice-level labels, multi-scale CNN activation maps, and candidate screening for fully automated nodule segmentation.On LIDC-IDRI, it achieves competitive performance compared with a CNN-based fully supervised segmentation method.
2 Method
The method adapts a classification CNN to generate nodule activation maps and uses them to guide coarse-to-fine candidate segmentation. Multi-scale localization narrows the search scope, while residual activation maps identify the candidate most responsible for the classification response.
- Nodule Activation Map: The CNN classifies CT slices and generates nodule activation maps from weighted convolutional activations.Its architecture combines a fully convolutional component, Conv+GAP structure, and final fully connected layer.
- Nodule Activation Map: Global average pooling summarizes convolution-unit activations, while fully connected weights determine their importance for nodule classification.Upsampling the weighted activation map identifies the discriminative image region relevant to nodules.
- Multi-GAP Segmentation: Multi-GAP CNNs combine activation maps from multiple scales to improve localization using shallower layers with higher spatial resolution.The concatenated GAP feature vectors feed the final fully connected layer.
- Coarse Segmentation: Nodule candidates are screened within a NAM-defined spatial scope and coarsely segmented with an ICM-based multi-phase method.The scope is defined by the most prominent watershed-processed NAM blob.
- Fine Segmentation: Residual NAMs identify the true candidate by masking each candidate and selecting the one causing the largest activation change within the screening scope.The current implementation targets one nodule per NAM.
- Multi-GAP Segmentation: The combined multi-GAP approach uses finer localization from multi-GAP NAMs and stronger discriminative power from one-GAP NAMs.A one-GAP model generates and compares residual NAMs when multi-GAP localization is insufficient.
3 Experimental Results
Experiments use the public LIDC-IDRI dataset and compare the weakly supervised framework with a fully supervised CNN using detection and segmentation metrics. The framework is also evaluated on slices containing two nodules.
- Data and setup: 1,010 thoracic CT scans from LIDC-IDRI were used, with lung-centered 384x384 axial slices and expert nodule delineations.
- Data and setup: Training, validation, and test subjects were separated in a 4:1:1 ratio using stratified sampling and non-overlapping subjects.
- Evaluation: The comparison reports detection TPR, two FPR measures, Dice, TP Dice, and TP DOA, with TP Dice and TP DOA additionally analyzed by nodule size.
- Models: The proposed network uses a VGG16-based architecture with one-, two-, or three-GAP configurations, while an adapted U-net provides the fully supervised comparison.
- Multiple nodules: For 108 two-nodule slices, the 2-GAP model detected both nodules in 50 slices and one nodule in another 42 slices.
4 Discussions and Conclusions
The paper presents a weakly supervised, multi-scale CNN framework for nodule segmentation using slice-level labels, reducing annotation demands while retaining competitive performance. Its remaining scope includes optional voxel supervision for contour refinement and future 3D extension.
- Contribution: The framework learns multi-scale discriminative regions from a classification CNN and uses candidate screening to localize and segment nodules with weak labels.
- Annotation efficiency: Annotation requirements are reduced by approximately 100 times relative to voxel-based labels, expressed as Nvoxel/Nslice ∼100.
- Performance: Detection compares very favorably with the fully supervised CNN, with higher TPR and much lower FPR, while segmentation approaches the benchmark for larger nodules.
- Performance: The fully supervised model remains more accurate on correctly detected nodules, especially around varied intensity patterns at edges, whereas the proposed method has smaller standard deviations.
- Future scope: Future extensions include combining NAM with patch-level or limited voxel-level labels and extending it to 3D CNNs for contextual information.