PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Segmentation of Pathological Lung Tissue with Dilated Fully Convolutional Networks

Marios Anthimopoulos, Stergios Christodoulidis, Lukas Ebner, Thomas Geiser, Andreas Christe, Stavroula Mougiakakou

arXiv:1803.06167v1cs.CV

TL;DR

ILD patterns in thoracic CT can be difficult to distinguish, motivating automated segmentation. The paper proposes an end-to-end, semi-supervised fully convolutional network using dilated convolutions to produce label maps. It reports performance surpassing prior studies with greater memory and computational efficiency.

  • Problem

    Similar manifestations of ILD pathologies in thoracic CT make their detection and recognition challenging, despite the importance of early diagnosis for treatment decisions.

    Source: §I · p. 1

  • Method

    A dilated fully convolutional network produces label maps from arbitrary-size HRCT images and is trained end-to-end using labeled and non-labeled regions.

    Source: §I · p. 2, B. Methods · p. 4

  • Results

    The proposed network surpassed the highest performance in previous studies and was more efficient in memory and computation; its accuracy was nearly equal to 82% with 58ms inference time.

    Source: B. Results · p. 6, §V · p. 7

  • Takeaways & Limitations

    Semi-supervised learning improved performance by nearly 1.5% without additional computational-resource requirements.

    Source: B. Results · p. 6

    The proposed work leaves the 3D nature of lung patterns and the bronchovascular tree for future consideration.

    Source: §V · p. 7

Abstract

from arXiv · show

Early and accurate diagnosis of interstitial lung diseases (ILDs) is crucial for making treatment decisions, but can be challenging even for experienced radiologists. The diagnostic procedure is based on the detection and recognition of the different ILD pathologies in thoracic CT scans, yet their manifestation often appears similar. In this study, we propose the use of a deep purely convolutional neural network for the semantic segmentation of ILD patterns, as the basic component of a computer aided diagnosis (CAD) system for ILDs. The proposed CNN, which consists of convolutional layers with dilated filters, takes as input a lung CT image of arbitrary size and outputs the corresponding label map. We trained and tested the network on a dataset of 172 sparsely annotated CT scans, within a cross-validation scheme. The training was performed in an end-to-end and semi-supervised fashion, utilizing both labeled and non-labeled image regions. The experimental results show significant performance improvement with respect to the state of the art.

I. INTRODUCTION

ILD diagnosis relies on recognizing pathological patterns in thoracic CT, but similar manifestations and the time demands of local sliding-window classification complicate the task. The paper proposes end-to-end, semi-supervised fully convolutional segmentation to produce label maps from arbitrary-size HRCT images.

I. INTRODUCTION

  • ILD diagnosis assesses pathological patterns in HRCT, but similar manifestations can make recognition challenging and misdiagnosis may have life-threatening consequences.CAD systems aim to automatically detect and recognize pathological lung tissue, which appears as textural patterns in CT scans.Source: §I · p. 1
  • Earlier CNN work for ILD used local sliding-window classification, which can be time-consuming and may miss useful less-local information.Medical-image training is also constrained by scarce data and the difficulty and time required to collect and annotate it.Source: §I · p. 1
  • The paper situates its approach against earlier local texture-classification work, including a prior CNN that outperformed traditional methods.A separate transfer-learning study using general texture databases yielded significant performance improvements.Source: §I · p. 1
  • The proposed deep fully convolutional network uses dilated convolutions and end-to-end semi-supervised training for ILD pattern recognition.It accepts an HRCT image of arbitrary size and outputs a corresponding label map, while using non-labeled regions during learning.Source: §I · p. 2

II. RELATED WORK

This section briefly surveys deep learning advances in computer vision and prior studies of ILD pattern classification.

II. RELATED WORK

  • The section reviews recent advances in deep learning for computer vision.The review is described as short.Source: §II · p. 2
  • It also provides a brief overview of earlier studies on ILD pattern classification.Source: §II · p. 2
  • The section pairs a computer-vision review with an overview of prior ILD pattern-classification studies.Source: §II · p. 2

A. Deep CNNs for Computer Vision

Deep CNN development has been accompanied by architectural changes that support deeper models and semantic segmentation. Fully convolutional networks preserve pixel-level spatial correspondence, while dilated convolutions enlarge receptive fields without downsampling feature maps.

A. Deep CNNs for Computer Vision

  • CNN advances included smaller 3×3 kernels in VGG, parallel operations in GoogleNet, and skip connections in ResNet that enabled training networks with hundreds of layers [9].Batch normalization also supported these developments by regularizing and accelerating training.Source: A. Deep CNNs for Computer Vision · p. 2
  • Semantic segmentation assigns a class label to every image pixel, and fully convolutional networks use spatially aligned feature maps to classify pixels for inputs of any size.Convolutional layers of size 1×1 can serve as local dense layers.Source: A. Deep CNNs for Computer Vision · p. 2
  • Encoder-decoder architectures and upsampling or transposed convolution restore input resolution after downsampling, while max-unpooling and skip-connected variants offer alternatives [15].Source: A. Deep CNNs for Computer Vision · p. 2
  • Dilated convolution inserts zero-valued gaps between kernel values, increasing receptive field without increasing parameter count or downsampling feature maps.Figure 1 illustrates kernels with different dilation rates; prior approaches use dilation for multiscale context [21].Source: A. Deep CNNs for Computer Vision · p. 2, Figure 1 · p. 2

B. ILD Pattern Classification

ILD pattern recognition has commonly been treated as local texture classification, first with handcrafted features and later with CNNs. Many CNN approaches remain patch-wise, while some pretrained architectures require image resizing and artificial three-channel inputs.

B. ILD Pattern Classification

  • Traditional ILD pattern classifiers commonly used handcrafted texture features with machine-learning classifiers in a local sliding-window framework.Examples include gray-level, co-occurrence, run-length and fractal features, filter responses, LBP, local DCT, bag-of-features and sparse-representation methods –.Source: B. ILD Pattern Classification · p. 3
  • Many CNN-based lung-pattern methods remain patch-wise, predicting class probabilities for each square patch –.Earlier studies used shallow networks, while prior work also explored deeper CNNs and transfer learning.Source: B. ILD Pattern Classification · p. 3
  • Pretrained AlexNet and GoogleNet approaches were adapted for slice-level pathology presence or absence, requiring image resizing and artificial three-channel inputs [34].Source: B. ILD Pattern Classification · p. 3

III. MATERIALS AND METHODS

This section introduces the materials used for training and testing, then presents the proposed network; Table I reports dataset statistics across six tissue classes.

III. MATERIALS AND METHODS

  • The section covers both the training and testing materials and the proposed segmentation network.The dataset statistics are organized by healthy tissue, ground glass opacity, micronodules, consolidation, reticulation, and honeycombing.Source: §III · p. 3, Table I · p. 3
  • Table I summarizes data statistics across the six considered tissue classes.The caption identifies the classes but does not provide their individual statistics.Source: Table I · p. 3

A. Materials

The study uses 172 HRCT scans from two hospitals, with radiologist annotations of typical ILD patterns and healthy tissue. The annotations are sparse and the classes are unevenly distributed.

A. Materials

  • The dataset contains 172 HRCT scans: 109 from Geneva University Hospital and 63 from Inselspital Bern.The scans were acquired between 2003 and 2015 using different scanners and protocols; HUG scans have 10–15 mm spacing, and slice thickness is 1–2 mm in both datasets.Source: A. Materials · p. 3
  • Two experienced radiologists annotated typical ILD patterns and healthy tissue, considering six tissue classes.The classes are normal, ground glass opacity, micronodules, consolidation, reticulation, and honeycombing; each case also has a lung-field segmentation mask.Source: A. Materials · p. 3
  • Annotations cover only typical manifestations, leaving other lung regions unlabeled and excluded from supervised training and evaluation.Sparse annotation permits more scans to be annotated for the same effort, increasing data diversity.Source: A. Materials · p. 3
  • The considered tissue classes are unevenly distributed across cases.The passage points to Table I for dataset statistics but does not provide the class counts.Source: A. Materials · p. 3

B. Methods

The method uses a dilated, purely convolutional network with dense feature aggregation and a semi-supervised loss. Training also uses weighted metrics and online image augmentation.

B. Methods

  • The 13-layer network uses dilated convolutions without downsampling, reaching a 287×287 receptive field.The first ten layers each have 32 3×3 kernels, with dilation rates 1, 1, 2, 3, 5, 8, 13, 21, 34, and 55; the rates mitigate gridding through denser sampling.Source: B. Methods · pp. 3–4, Figure 3 · p. 5, Figure 4 · p. 5
  • ITK-Snap was used for the annotation process.Source: B. Methods · p. 3
  • Concatenating the input with outputs from the first ten layers yields 321 feature maps, which feed 1×1 layers that produce six-class pixel probabilities.The concatenation aggregates features across scales and abstraction levels and facilitates gradient flow; dropout is applied at rate 0.5.Source: B. Methods · p. 4
  • Instance normalization and a skip connection combine normalized and non-normalized activations before ReLU.The connection diminishes mean normalization and provides feature-contrast enhancement without complete invariance to variance shifts.Source: B. Methods · p. 4
  • Training minimizes class-weighted cross-entropy plus an entropy term on unlabeled regions, with the unsupervised contribution controlled by a weight.Class weights are inversely proportional to class sample counts, so classes contribute equally to the weighted metrics.Source: B. Methods · p. 4
  • Training stops after 50 epochs without significant validation improvement, and online augmentation randomly applies one of eight flip-and-rotation combinations per image per epoch.The augmentation is intended to increase training data and avoid overfitting.Source: B. Methods · pp. 4–5

IV. EXPERIMENTAL SETUP AND RESULTS

The section presents the experimental setup and results, using them to justify the proposed method’s algorithmic choices and compare it with the state of the art.

IV. EXPERIMENTAL SETUP AND RESULTS

  • The experiments are intended to justify the method’s algorithmic choices and compare it with the state of the art.Source: §IV · p. 5

A. Experimental Setup

The study evaluated its models using case-level five-fold cross-validation on 172 scans, with balanced accuracy measured over annotated regions. The setup also balanced class distributions across folds and limited inputs to lung crops with annotation context.

A. Experimental Setup

  • The 172 scans were split into five non-overlapping, case-level folds, with each test fold held out while the remaining scans trained the model.One fold contained 36 scans and the other four contained 34 each.Source: A. Experimental Setup · p. 5
  • Balanced accuracy was averaged across five folds and calculated only where ground-truth annotations were available.Source: A. Experimental Setup · p. 5
  • Fold assignment used hill climbing to maximize class-distribution entropy by swapping cases between sets when this improved balance.Source: A. Experimental Setup · p. 5
  • The networks received cropped left and right lung regions from slices with relevant annotations, with a 32-pixel margin added on each side.Source: A. Experimental Setup · pp. 5–6
  • Experiments used Python2, Keras, and Theano on Linux with an Intel Core i7-5960X CPU, NVIDIA GeForce Titan X GPU, and 128 GB RAM.Source: A. Experimental Setup · p. 6

B. Results

The proposed network outperformed previous studies while using less memory and computation. Configuration experiments found that semi-supervised learning improved accuracy without added computational requirements, and that normalization, dilation, and the selected unsupervised weight affected performance.

B. Results

  • Nearly 1.5% improvement came from semi-supervised learning, with no additional computational requirements.Source: B. Results · p. 6
  • Accuracy was nearly 82% at 58 ms inference time; 64 kernels improved results nonsignificantly, while 16 kernels performed notably worse.Source: B. Results · p. 6
  • Accuracy decreased almost 3% with exponential dilation rates, while non-dilated convolutions performed poorly after the receptive field shrank.Source: B. Results · p. 6
  • The proposed network yielded the best results while running faster and using far fewer parameters than previous studies.Source: Table III · p. 6, B. Results · p. 7
  • Adding the InstanceNorm skip connection further improved accuracy, while omitting normalization or concatenation of the first 10 layers impaired results.Source: B. Results · p. 6
  • The best-performing unsupervised weight was wu = 0.1, which was used to train the proposed model.Source: B. Results · p. 6

V. CONCLUSIONS

The study concludes that its fully convolutional, dilated CNN can segment pathological lung tissue on HRCT slices and surpass previous studies with greater memory and computational efficiency. The 172-case evaluation used cross-validation, while future work includes modeling 3D patterns and the bronchovascular tree.

V. CONCLUSIONS

  • The fully convolutional CNN surpassed previous studies and was more efficient in memory and computation.Source: §V · p. 7
  • Evaluation used cross-validation on 172 cases, with dataset folds split by case.Source: §V · p. 7
  • Future work includes extending the model to consider 3D lung patterns and account for the bronchovascular tree.Source: §V · p. 7

Related papers

Loading 1803.06167v1…