PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepPap: Deep Convolutional Networks for Cervical Cell Classification

Ling Zhang, Le Lu, Isabella Nogues, Ronald M. Summers, Shaoxiong Liu, Jianhua Yao

arXiv:1801.08616v1cs.CV

TL;DR

Automation-assisted cervical screening still faces inadequate performance, limited labeled data, and challenges associated with accurate segmentation. DeepPap uses transfer-learned ConvNets to classify coarsely nucleus-centered image patches without prior segmentation, achieving superior performance on cervical-cell datasets, including 98.3% accuracy and specificity on Herlev.

  • Problem

    Automation-assisted reading performance is not yet adequate for primary cervical screening, while cervical-cell datasets contain limited labeled data and accurate segmentation remains challenging.

  • Method

    DeepPap fine-tunes an ImageNet-pre-trained ConvNet on raw RGB image patches coarsely centered on nuclei and aggregates patch predictions to classify cells as abnormal or normal.

  • Results

    DeepPap generates the highest performances on both the Herlev Pap smear and HEMLBC datasets, with 98.3% accuracy and 98.3% specificity on Herlev.

  • Takeaways & Limitations

    The method automatically extracts deep features without requiring cytoplasm or nucleus segmentation and has potential for automation-assisted cervical screening systems.

  • Takeaways & Limitations

    Clinical adoption is constrained by slow inference and missed abnormal cells, including 4.1% of severe dysplasia and 2% of carcinoma cells misclassified as normal.

Abstract

from arXiv · show

Automation-assisted cervical screening via Pap smear or liquid-based cytology (LBC) is a highly effective cell imaging based cancer detection tool, where cells are partitioned into "abnormal" and "normal" categories. However, the success of most traditional classification methods relies on the presence of accurate cell segmentations. Despite sixty years of research in this field, accurate segmentation remains a challenge in the presence of cell clusters and pathologies. Moreover, previous classification methods are only built upon the extraction of hand-crafted features, such as morphology and texture. This paper addresses these limitations by proposing a method to directly classify cervical cells - without prior segmentation - based on deep features, using convolutional neural networks (ConvNets). First, the ConvNet is pre-trained on a natural image dataset. It is subsequently fine-tuned on a cervical cell dataset consisting of adaptively re-sampled image patches coarsely centered on the nuclei. In the testing phase, aggregation is used to average the prediction scores of a similar set of image patches. The proposed method is evaluated on both Pap smear and LBC datasets. Results show that our method outperforms previous algorithms in classification accuracy (98.3%), area under the curve (AUC) (0.99) values, and especially specificity (98.3%), when applied to the Herlev benchmark Pap smear dataset and evaluated using five-fold cross-validation. Similar superior performances are also achieved on the HEMLBC (H&E stained manual LBC) dataset. Our method is promising for the development of automation-assisted reading systems in primary cervical screening.

1 INTRODUCTION

Cervical screening systems remain constrained by difficult cell segmentation, limited labeled data, and reliance on handcrafted features. Deep ConvNets with transfer learning are proposed to classify cervical cells from coarsely centered patches, achieving strong results on two datasets.

  • Motivation: Accurate segmentation remains difficult because cell clusters and abnormal-versus-normal nuclear variation obstruct individual cytoplasm and nucleus delineation.Herlev nucleus segmentation accuracy ranges from 0.85 to 0.92.
  • Motivation: Current systems commonly use handcrafted morphology, chromatin, and texture features organized through feature selection or dimensionality reduction.These features follow cytology rules or prior computer-aided-diagnosis experience.
  • Motivation: Cervical-cell datasets are small because high-expertise annotation is required; Herlev contains only 917 cells.The dataset includes 675 abnormal and 242 normal cells.
  • Proposed approach: The proposed method classifies RGB patches coarsely centered on nuclei, using ImageNet pre-training, cervical-cell fine-tuning, and aggregation of patch predictions.It is evaluated on Herlev Pap smear and HEMLBC datasets using five-fold cross-validation.
  • Contribution: 98.3% accuracy is achieved on Herlev and 98.6% on HEMLBC, surpassing previous best accuracies of 96.8% and 94.3%, respectively.The method also avoids dependence on accurate segmentation and handcrafted features.

2 METHODS

The method has separate training and testing stages built around transfer learning and aggregated ConvNet outputs. A network is pre-trained on ImageNet, fine-tuned on preprocessed cervical-cell samples, and used to produce an abnormality score.

  • Training: Training pre-trains a ConvNet on ImageNet, preprocesses cervical-cell data, transfers network parameters, and fine-tunes the resulting model.The transferred parameters initialize the new ConvNet before training on cervical-cell samples.
  • Testing: Testing feeds preprocessed images into the fine-tuned ConvNet and aggregates its output values to obtain the abnormality score.Aggregation produces the final test-time scoring signal.

2.1 Data Preprocessing

Preprocessing avoids full-cell pre-segmentation by extracting augmented square patches around nuclei. Rotations and translations increase training variation while preserving nucleus size and intensity as discriminative cues.

  • Patch extraction: The method uses nucleus-centered patches rather than full-cell images because obtaining individual cells requires challenging pre-segmentation.Nucleus features are treated as substantially discriminative for cervical abnormalities.
  • Patch extraction: Rotating each cell image generates Nr patches centered on rotated nucleus centers, increasing the number of training samples.Cells are treated as rotationally invariant for this augmentation.
  • Patch extraction: Random translations of each nucleus centroid generate Nt patches that simulate inaccurate nucleus detection and enlarge the training set.Translations are limited to d pixels.
  • Augmentation constraints: Scale and color transformations are omitted because nucleus size and intensity are essential abnormality cues.The preprocessing therefore preserves these properties rather than transforming them.
  • Class balance: Class imbalance in Herlev creates a risk of majority-class bias, while excessive false positives on normal cells are undesirable in practice.Herlev contains approximately three times more abnormal than normal cell images.

2.2 Convolutional Neural Networks

The ConvNet maps raw image pixels through convolutional, nonlinear, pooling, and fully connected stages to classify cervical cells. Its parameters can be initialized from ImageNet and optimized with backpropagation-based training.

  • Architecture: A ConvNet consists of convolutional, nonlinearity, pooling, and fully connected layers that transform raw pixel intensities into class outputs.The cervical-cell model receives mean-subtracted images and has output neurons corresponding to classes.
  • Convolutional layer: Convolution applies learned filters to local rectangular patches, while ReLU nonlinearity uses f(x) = max(0, x).The paper states that ReLU increases training speed.
  • Pooling: Pooling downsamples feature maps by summarizing local responses and provides invariance to minor translations.Max-pooling is given as a common implementation.
  • Fully connected layers: The first fully connected layers fuse feature maps into a feature vector, while the final layer computes class probabilities with softmax regression.Dropout is used in fully connected layers to reduce overfitting.
  • Optimization: The training procedure optimizes ConvNet weights by minimizing classification error with backpropagation and stochastic gradient descent.Momentum and weight decay are used during training to speed learning and reduce overfitting.

2.3 Transfer Learning

The method transfers early convolutional and pooling layers from an ImageNet-pre-trained ConvNet and fine-tunes them with cervical-cell data. Task-specific fully connected layers are added and trained for abnormal/normal classification.

  • ImageNet-pre-trained convolutional and pooling layers are transferred into the cervical-cell network and jointly fine-tuned on the target dataset.The transferred layers use a learning rate 10 times smaller than the newly initialized fully connected layers.
  • The transferred base uses the same conv and pool layer locations as BVLC CaffeNet, while task-specific fully connected layers are randomly initialized.The network accepts RGB channels, matching the transferred CaffeNet input format.

2.4 Testing

Testing generates multiple transformed views and crops around each nucleus, then aggregates their ConvNet predictions into a final cell score. The procedure uses the same 917-cell Herlev dataset described in Table 1.

  • For each unseen cell, random-view aggregation generates Nv rotated and translated patches around the nucleus centroid.Each augmented patch contributes multiple corner, center, and mirrored sub-crops.
  • The Herlev benchmark contains 917 cells, including 242 normal and 675 abnormal cells.

3 EXPERIMENTAL METHODS

The experiments use Pap-smear and H&E-stained manual LBC datasets, train an ImageNet-initialized ConvNet with augmented patches, and evaluate it through five-fold cross-validation and multiple classification analyses.

  • Datasets: The study evaluates cervical-cell classification on two datasets acquired with different slide preparation, staining, and imaging conditions.The datasets represent conventional Pap smear and manual liquid-based cytology with H&E staining.
  • Datasets: The Herlev dataset contains 917 single-cell images with ground-truth segmentation and classification, including seven diagnostic classes.Its images were prepared using conventional Pap smear and Pap staining at 0.201 µm per pixel.
  • Training and testing protocols: Training augmentation applies rotations and translations to create 100 patches per abnormal Herlev cell and 280 per normal cell, balancing the training distribution.The augmentation settings differ between abnormal and normal cells only during training, not testing.
  • Datasets: The HEMLBC dataset includes 989 abnormal and 989 randomly selected normal cells, with abnormal diagnoses confirmed by two experienced pathologists.The abnormal cells come from eight biopsy-confirmed CIN slides and the normal cells from eight NILM slides.
  • Network architectures and implementation: ConvNet-T transfers five ImageNet-trained weight layers and replaces the original classifier with fully connected layers sized 1024-256-2 for binary classification.The transferred network uses conv1 through pool5, while fc6 through fc8 are randomly initialized.
  • Evaluation: Performance is estimated by averaging results across five validation folds after splitting the cell population before data augmentation.The evaluation also examines ROC-based scores, robustness to nucleus-center localization error, and a seven-class classification variant.

4 RESULTS

The ConvNet learns useful cervical-cell representations and achieves strong classification performance on Herlev, with improved results also observed on HEMLBC.

  • ConvNet learning results: After six epochs, validation loss reached 0.119 with validation accuracy of 0.972 during Herlev fine-tuning.The training and validation process was shown over 30 epochs.
  • Quantitative results: 98.2% sensitivity, 98.3% specificity, 98.3% accuracy, 98.3% H-mean, 98.8% F-measure, and 0.998 AUC were achieved on Herlev.The method outperformed previous methods on all metrics except sensitivity, which was slightly lower.
  • Quantitative results: 98.3% specificity exceeded the previous highest result of 92.2% on Herlev.
  • Quantitative results: The method showed perfect performance for two normal cell classes and mild dysplasia, while columnar epithelial and severe dysplasia reached 95.9%.Most abnormality scores were close to 0 or 1, with some misclassifications involving normal columnar and severe squamous cells.
  • Quantitative results: 1.6% average overall error was achieved for seven-class classification, compared with 7.9% and 3.9% for previous methods.
  • HEMLBC results: The method's performance improved over the previous MLP-based method on the slightly smaller HEMLBC dataset.

5 DISCUSSION

The proposed ConvNet classifies cervical cells from coarsely centered image patches without segmentation and remains robust to inaccurate nucleus centroids. It distinguishes difficult cases and achieves high sensitivity and specificity across Herlev and HEMLBC, while retaining important clinical limitations.

  • Qualitative results: The discussion illustrates correct classifications across four abnormal Herlev classes and three normal classes using abnormal scores.A score of 1 represents a 100% probability of abnormality.
  • Method robustness: The method requires only a coarse nucleus centroid, avoiding cytoplasm and nucleus segmentation.Its robustness to inaccurate centroid detection is evaluated through ConvNet-T∆5 and ∆10.
  • Dataset performance: The method produces the highest performances on both the Herlev Pap smear and HEMLBC liquid-based cytology datasets.The reported advantage includes high sensitivity and especially high specificity.
  • Limitations: Classification requires 3.5 seconds per patch, although removing 100-variant test-time augmentation reduces this to 0.035 seconds with approximately 1% accuracy loss.The method also misclassifies some severe dysplasia and carcinoma cells as normal, while both datasets mainly contain expert-selected typical cells.

6 CONCLUSION

The paper proposes a convolutional neural network that classifies cervical cells from image patches coarsely centered on nuclei, without relying on segmentation or hand-crafted features. Transfer learning, fine-tuning, and prediction aggregation produce the highest reported performance on both Herlev and HEMLBC datasets.

  • 6 CONCLUSION: The method automatically extracts deep features from cervical cell image patches instead of using segmented cells and hand-crafted features.Its inputs are patches coarsely centered on the nucleus.
  • 6 CONCLUSION: The pipeline transfers features from a pre-trained model, fine-tunes a new ConvNet on cell patches, and aggregates multiple predictions.Aggregation forms the final network output.
  • 6 CONCLUSION: The proposed method yields the highest performance on both the Herlev Pap smear and HEMLBC liquid-based cytology datasets compared with previous methods.
Loading 1801.08616v1…