PaperlayerView exact revision on arXiv

Source-linked AI summary

Computer-aided diagnosis in histopathological images of the endometrium using a convolutional neural network and attention mechanisms

Hao Sun, Xianxu Zeng, Tao Xu, Gang Peng, Yutao Ma

arXiv:1904.10626v1cs.CV

TL;DR

Endometrial CADx systems have faced limits from small training datasets and handcrafted features, despite histopathology being the diagnostic gold standard. HIENet combines a CNN with attention mechanisms to classify H&E images and highlight morphology-related features. It performed well in cross-validation and external validation, with the authors reporting better classification accuracies than three pathologists.

  • Problem

    Prior endometrial CADx approaches were limited by small training samples and handcrafted features, while histopathology is the diagnostic gold standard.

    Source: Introduction · pp. 3–4

  • Method

    HIENet applies a CNN with attention mechanisms and feature-map visualization to classify H&E endometrial images and highlight morphology-related features.

    Source: Introduction · p. 4, Conclusion · p. 18

  • Results

    HIENet achieved 84.50% accuracy in four-class external validation, and the authors report better classification accuracies than three associate chief physicians.

    Source: External validation result · p. 15, Conclusion · p. 18

  • Takeaways & Limitations

    HIENet can highlight morphological characteristics in H&E images to provide pathologists with pixel-level feature interpretability.

    Source: Conclusion · p. 18

    HIENet had a higher false negative rate than the three human experts, with some EA images misclassified as EH amid insufficient EA and complex-EH images and low-magnification inputs.

    Source: Discussion · p. 16

Abstract

from arXiv · show

Uterine cancer, also known as endometrial cancer, can seriously affect the female reproductive organs, and histopathological image analysis is the gold standard for diagnosing endometrial cancer. However, due to the limited capability of modeling the complicated relationships between histopathological images and their interpretations, these computer-aided diagnosis (CADx) approaches based on traditional machine learning algorithms often failed to achieve satisfying results. In this study, we developed a CADx approach using a convolutional neural network (CNN) and attention mechanisms, called HIENet. Because HIENet used the attention mechanisms and feature map visualization techniques, it can provide pathologists better interpretability of diagnoses by highlighting the histopathological correlations of local (pixel-level) image features to morphological characteristics of endometrial tissue. In the ten-fold cross-validation process, the CADx approach, HIENet, achieved a 76.91 $\pm$ 1.17% (mean $\pm$ s. d.) classification accuracy for four classes of endometrial tissue, namely normal endometrium, endometrial polyp, endometrial hyperplasia, and endometrial adenocarcinoma. Also, HIENet achieved an area-under-the-curve (AUC) of 0.9579 $\pm$ 0.0103 with an 81.04 $\pm$ 3.87% sensitivity and 94.78 $\pm$ 0.87% specificity in a binary classification task that detected endometrioid adenocarcinoma (Malignant). Besides, in the external validation process, HIENet achieved an 84.50% accuracy in the four-class classification task, and it achieved an AUC of 0.9829 with a 77.97% (95% CI, 65.27%-87.71%) sensitivity and 100% (95% CI, 97.42%-100.00%) specificity. In summary, the proposed CADx approach, HIENet, outperformed three human experts and four end-to-end CNN-based classifiers on this small-scale dataset composed of 3,500 hematoxylin and eosin (H&E) images regarding overall classification performance.

Introduction

Endometrial CADx systems have been limited by small training samples and reliance on handcrafted features, while histopathology remains the diagnostic gold standard. This study aims to develop a deep-learning approach that assists pathologists with H&E image classification and interpretable tissue-feature correlations.

Introduction

  • Endometrial cancer is a major health concern; the paper reports 319,600 new uterine-corpus cancer cases worldwide in 2012 and rising endometrial-cancer incidence.Appropriate treatment after diagnosis is associated in the passage with a five-year survival rate over 80%.Source: Introduction · p. 3
  • Existing endometrial CADx systems were limited by small training samples and handcrafted features; one SVM achieved 81% accuracy on 516 regions of interest.Histopathological confirmation is the gold standard for diagnosing endometrial cancer.Source: Introduction · pp. 3–4
  • Digital pathology is presented as a potential response to pathologist shortages and the pressure for rapid review, which can yield conflicting or incorrect reports.The passage describes this need particularly in developing countries.Source: Introduction · p. 4
  • The study develops deep-learning CADx to classify H&E endometrial tissue images and provide pathologists with interpretable correlations between image features and tissue morphology.The authors expect this assistance may improve pathologists’ efficiency and productivity.Source: Introduction · p. 4

Materials and Methods

The study collected endometrial specimens, prepared and stained tissue sections, digitized selected slides, and extracted labeled histopathological images for analysis.

Sample collection

  • 498 endometrial tissue samples were collected from patients treated at the study hospital between October 2017 and August 2018.Patients had not received hormone therapy, radiotherapy, or chemotherapy before surgery; personal information was not collected.Source: Sample collection · p. 4

Sample preparation

  • Specimens were formalin-fixed, dehydrated, paraffin-embedded, sectioned into 4 μm slices, and stained with H&E for pathologist review.Three pathologists with more than ten years of experience selected representative slides with unanimous diagnoses.Source: Sample preparation · pp. 4–5

Digital imaging

  • Selected H&E slides were scanned at 10x or 20x magnification and saved as JPEG images.Source: Digital imaging · p. 5

Histopathological image processing

Taxonomy of digital histopathological image labels

The dataset defines four endometrial tissue classes and two broader benign/malignant classes. Four end-to-end CNN classifiers were trained as performance comparators.

Taxonomy of digital histopathological image labels

  • The dataset contains 3,302 labeled JPEG histopathological images, with NE, EP, EH, and EA as four classes and EA designated malignant versus the others benign.NE and EH each have subtypes; the passage lists 600, 21, and 712 NE images and 516 simple and 282 complex EH images, alongside 636 EP and 535 EA images.Source: Taxonomy of digital histopathological… · pp. 5–6

End-to-end convolutional neural networks for comparison

  • Training reduced the learning rate when classification accuracy on training data failed to improve for p epochs.Here, p denotes the number of training epochs in the patience setting.Source: End-to-end convolutional neural… · p. 6
  • Four comparison classifiers used AlexNet, VGG-16, InceptionV3, and ResNet-50, fine-tuned from pretrained ImageNet weights.The classifiers were trained end to end with different hyperparameters.Source: End-to-end convolutional neural… · p. 6

HIENet

HIENet combines a VGG-16 backbone with position and channel attention blocks to process endometrial histopathology images. The blocks respectively capture relationships among local areas and learn channel-wise feature weights.

HIENet

  • HIENet uses a VGG-16 backbone and concatenates position-attention, channel-attention, and original feature maps before pooling and classification.The figure depicts the overall architecture and the two attention-block components.Source: HIENet · p. 6, Figure 1 · p. 7
  • Position Attention applies non-local self-attention to capture context relations between different local image areas.Its convolutional layers learn context relations in parallel before producing a new feature map.Source: HIENet · p. 7
  • Channel Attention learns weights for feature-map channels and combines the new weights with the original features to generate a feature map.The block initializes channel weights with global average pooling and learns new weights using two fully connected layers.Source: HIENet · pp. 7–8

Image processing for model training

The classifiers used normalized, augmented whole-image inputs, and evaluation combined ten-fold cross-validation with a blinded comparison against pathologists and several classification metrics.

Image processing for model training

  • Whole images were resized, z-score normalized, and randomly flipped during training, with batch normalization applied to layer inputs.Inputs were 224 x 224 pixels, except InceptionV3 inputs at 299 x 299 pixels; horizontal and vertical flips each had a 50% probability.Source: Image processing for model training · p. 8

Evaluation methods

  • Ten-fold cross-validation repeatedly held out one subsample for testing and averaged results across all ten test folds.Each fold trained on the remaining nine subsamples.Source: Evaluation methods · p. 8
  • Three experienced pathologists completed a blinded test on 100 label-free images after repeated pre-testing with ground-truth feedback.Their benign-versus-malignant likelihoods were formed by summing probabilities over subclasses.Source: Evaluation methods · p. 8

Evaluation metrics

  • Accuracy, sensitivity, and specificity quantified classification, while ROC curves and AUC assessed binary classification performance.ROC curves plot sensitivity against 1 – specificity across thresholds; higher AUC indicates better binary classification results.Source: Evaluation metrics · pp. 8–9

Image feature visualization

HIENet’s feature maps were visualized to relate image features to histopathological interpretations, and a new patient-derived image dataset was collected for external validation. The study’s processing and classification code is publicly available.

Image feature visualization

  • Guided backpropagation and class activation maps visualize HIENet features, with CAM heat maps highlighting image regions relevant to a given class.The approach relates pixel-level image features to histopathological interpretations.Source: Image feature visualization · p. 9

External Validation

  • The external-validation dataset contains 200 images from 50 randomly selected patients examined at the study hospital in the first quarter of 2019.It includes 74 normal endometrium, 12 endometrial polyp, 55 endometrial hyperplasia, and 59 endometrial adenocarcinoma images.Source: External Validation · p. 9

Code availability

  • The image-processing and classification code is available in the study’s public repository.The experiments used Python 3.6.5, Keras 2.1.3, and TensorFlow 1.2.0.Source: Code availability · p. 10

Results

HIENet performed well in comparisons with CNN classifiers and human investigators, and its maps highlighted tissue morphology across endometrial classes. External validation also showed strong classification results.

Comparison among different CNN-based classifiers

Comparison between HIENet and human experts

Visualization of pixel-level morphological features

  • HIENet's maps highlighted serrated or regular glands in normal tissue, polyp vessels, hyperplastic gland dilation or crowding, and adenocarcinoma back-to-back glands.The visualizations included H&E images and learned pixel-level GB and CAM maps.Source: Figure 6 · p. 14, Visualization of pixel-level… · pp. 14–15

External validation result

Discussion

HIENet’s errors were concentrated in difficult endometrial adenocarcinoma (EA) cases, with limited data and low-magnification images constraining discrimination from hyperplasia. The authors describe a potential human–machine workflow and outline data and model-development priorities, while noting that direct clinical use is not yet possible.

Discussion

  • The 3,500-image dataset was small, and the authors identify scarce labeled medical images as the primary challenge in endometrial image analysis.They plan to collect more tissue samples, particularly endometrial cancer images.Source: Discussion · p. 16
  • HIENet had a relatively higher false-negative rate than the three human experts, partly because EA and Complex EH examples were insufficient and subtle cellular cues were hard to detect.Images were low magnification, and HIENet did not segment them according to morphological characteristics.Source: Discussion · p. 16
  • HIENet misclassified three EA cases as EH, although all three investigators diagnosed them correctly.The cases had irregular glands scattered disorderly, and HIENet highlighted regions different from those recognized by investigators.Source: Figure 8 · p. 17, Discussion · p. 17
  • Although HIENet cannot be applied directly in clinical environments, the authors propose human–machine collaboration in which HIENetBC screens negative samples and flags suspected lesion areas for pathologists.In external validation, HIENetBC achieved an AUC of 0.9971, with 86.44% sensitivity and 100% specificity.Source: Discussion · p. 17
  • Future work includes adding data types such as atypical endometrial hyperplasia, building an expert-in-the-loop framework, and applying reinforcement or transfer learning.The dataset currently contains no images of atypical endometrial hyperplasia.Source: Discussion · pp. 17–18

Conclusion

HIENet combines a CNN with attention mechanisms for endometrial histopathology classification and highlights morphological characteristics in H&E images. The authors report classification performance on a small dataset and identify potential for human–machine collaboration in grading diagnosis.

Conclusion

  • HIENet was effective in binary and multiclass classification on 3,500 H&E images across cross-validation and external validation, with better classification accuracies than three associate chief physicians.The authors describe the dataset as small-scale.Source: Conclusion · p. 18
  • Attention mechanisms and a class activation map highlight morphological characteristics in H&E images to provide pathologists with pixel-level histopathological correlations.Source: Conclusion · p. 18
  • HIENet may have potential in human–machine collaboration for grading endometrial diseases and could help increase pathologists’ productivity.Source: Conclusion · p. 18

Data citation

The cited dataset is a 2018 figshare resource by Zeng, Sun, and Ma.

Data citation

  • Zeng, Sun, and Ma’s cited resource is titled “A histopathological image dataset for endometrial disease diagnosis.”Source: Data citation · p. 22
  • The dataset citation identifies figshare as the repository and gives DOI 10.6084/m9.figshare.7306361.v2.Source: Data citation · p. 22
  • The cited dataset record is dated 2018.Source: Data citation · p. 22

Additional Information

The additional information includes a position-attention feature-map operation, an appendix confusion-matrix reference, and a competing-interests statement.

Additional Information

  • The authors declare no competing financial interests.Source: Additional Information · p. 22
  • The Position Attention block takes feature map F as input, where w, h, and c denote its width, height, and channel dimensions.Each feature contains local information from the input image.Source: Appendix S2 · p. 24
  • The position-attention operation forms K and V through convolution and reshaping, with max pooling in the V branch, then uses feature correlations to obtain position-attention weights and generate a context-informed feature map.The output feature map has the same size as F; the stated final operations include 1x1 convolution and batch normalization.Source: Appendix S2 · p. 24
  • Appendix S4 provides HIENetBC confusion matrices for two-class ten-fold cross-validation and external validation.Source: Appendix S2 · p. 26

Related papers

Loading 1904.10626v1…