Source-linked AI summary
DeepCervix: A Deep Learning-based Framework for the Classification of Cervical Cells Using Hybrid Deep Feature Fusion Techniques
Md Mamunur Rahaman, Chen Li, Yudong Yao, Frank Kulwa, Xiangchen Wu, Xiaoyan Li, Qian Wang
TL;DR
Manual cervical-cell screening is time-consuming and error-prone, while existing automated methods face segmentation, handcrafted-feature, and imbalanced multiclass challenges. DeepCervix addresses these issues with hybrid deep feature fusion and achieves high accuracy across SIPAKMED and Herlev classification tasks.
Problem
Existing cervical-cell classification methods often require difficult pre-segmentation, rely on handcrafted features, or perform poorly on imbalanced multiclass data, limiting automated screening support.
Method
DeepCervix fine-tunes pretrained VGG16, VGG19, XceptionNet, and ResNet50 models and adaptively fuses their complementary deep features for end-to-end cervical-cell classification.
Results
The method achieves 99.85%, 99.38%, and 99.14% accuracy on SIPAKMED 2-class, 3-class, and 5-class tasks, plus 98.91% and 90.32% on Herlev binary and 7-class tasks.
Takeaways & Limitations
DeepCervix provides end-to-end cervical-cell classification without pre-segmentation or handcrafted features and shows potential for improved cervical cancer diagnostic systems.
Takeaways & Limitations
Performance degrades for 7-class classification on Herlev, and the method is not yet generalized to classification involving overlapping cells.
Abstract
from arXiv · showhide
Cervical cancer, one of the most common fatal cancers among women, can be prevented by regular screening to detect any precancerous lesions at early stages and treat them. Pap smear test is a widely performed screening technique for early detection of cervical cancer, whereas this manual screening method suffers from high false-positive results because of human errors. To improve the manual screening practice, machine learning (ML) and deep learning (DL) based computer-aided diagnostic (CAD) systems have been investigated widely to classify cervical pap cells. Most of the existing researches require pre-segmented images to obtain good classification results, whereas accurate cervical cell segmentation is challenging because of cell clustering. Some studies rely on handcrafted features, which cannot guarantee the classification stage's optimality. Moreover, DL provides poor performance for a multiclass classification task when there is an uneven distribution of data, which is prevalent in the cervical cell dataset. This investigation has addressed those limitations by proposing DeepCervix, a hybrid deep feature fusion (HDFF) technique based on DL to classify the cervical cells accurately. Our proposed method uses various DL models to capture more potential information to enhance classification performance. Our proposed HDFF method is tested on the publicly available SIPAKMED dataset and compared the performance with base DL models and the LF method. For the SIPAKMED dataset, we have obtained the state-of-the-art classification accuracy of 99.85%, 99.38%, and 99.14% for 2-class, 3-class, and 5-class classification. Moreover, our method is tested on the Herlev dataset and achieves an accuracy of 98.32% for binary class and 90.32% for 7-class classification.
2 Introduction
DeepCervix addresses limitations in cervical-cell CAD by combining pretrained deep models and hybrid feature fusion for end-to-end multiclass classification without requiring pre-segmented images.
- Motivation: Manual Pap smear screening is time-consuming and error-prone, motivating computer-aided cervical-cell classification.Slides may contain around three million cells, making expert analysis difficult and tedious.
- Research gap: Existing CAD pipelines commonly involve segmentation and handcrafted features, while uneven multiclass data can reduce deep-learning performance.Cell clustering makes accurate segmentation challenging, and handcrafted features cannot ensure optimal classification.
- Proposed approach: DeepCervix uses pretrained ImageNet models, two-stage augmentation, and hybrid deep feature fusion to classify cervical cytopathology cells.The framework combines VGG16, VGG19, XceptionNet, and ResNet50 features through an improved fusion network.
- Workflow: The workflow retrieves cervical-cell images, augments them, processes them with four deep-learning models, fuses features, and evaluates precision, recall, F1 score, and accuracy.Unseen test images are supplied for classification after training.
- Contributions: The paper presents HDFF as its first reported application to cervical cytopathology-cell classification and reports the highest SIPAKMED accuracy.The authors also identify an improved fusion network combining dense, SoftMax, batch-normalization, and dropout layers.
3 Literature Review
Prior cervical-cell studies use diverse handcrafted, deep-learning, transfer-learning, and fusion methods, but the literature is weighted toward binary classification and often depends on segmentation or traditional features.
- Cervical-cell classification: Existing cervical-cell studies have used handcrafted histogram, texture, gray-level, and local-binary-pattern features with hybrid classifiers.Reported classifiers include support vector machines and adaptive neuro-fuzzy systems.
- Deep learning approaches: Deep CNN and transfer-learning approaches have been applied with feature selection, patch extraction, and classifiers including LSSVM and SoftMax regression.Some approaches use AlexNet or VGG16-based feature extraction for cervical-cell analysis.
- Segmentation requirements: Several deep models require pre-segmented cytoplasm or nucleus images, although one AlexNet-based approach reports classification without segmentation.The reviewed architectures include AlexNet, GoogleNet, ResNet, DenseNet, and VGG-like networks.
- Identified gap: The review identifies a literature emphasis on binary classification, limitations in transferred models’ medical-image representations, and uncertainty about traditional-feature optimality.These observations motivate methods addressing multiclass cervical-cell classification and feature representation.
- Feature fusion: Feature-fusion research has combined handcrafted and deep features or fused representations from multiple deep architectures in medical and remote-sensing imaging.Examples include glaucoma diagnosis and multi-structure scene classification.
4 Method
The method combines augmentation, transfer learning, multiple CNN architectures, late fusion, and hybrid deep feature fusion to classify cervical cells. Pretrained networks are adapted to cervical-cell images, and their complementary feature vectors are integrated for classification.
- Preprocessing: Data augmentation applies geometric transformations, image-processing functions, and random in-place image reconstruction during training.The transformations include affine operations, noise, edge and contrast processing, brightness changes, and flips.
- Deep learning: CNN-based classification uses convolutional, pooling, and fully connected layers to extract progressively higher-level image features.The study evaluates VGG16, VGG19, ResNet50, and XceptionNet architectures.
- Deep learning: ResNet uses skip connections so the network learns a residual function alongside the original input.The residual block is represented as H(x) = F(x) + x.
- Transfer learning: Transfer learning starts from pretrained models and fine-tunes them on cervical-cell images instead of training from scratch.The study freezes earlier layers and retrains later layers to combine generic and dataset-specific features.
- Fusion methods: Late fusion combines decisions from VGG16, VGG19, ResNet50, and XceptionNet using majority voting.Each class is selected according to the highest number of classifier votes.
- Hybrid deep feature fusion: DeepCervix integrates normalized feature vectors from four CNN architectures through concatenation and a softmax classification layer.The fused vectors are extracted from dense layers with 1024 neurons in VGG16, VGG19, ResNet50, and XceptionNet.
5 Experiments and Analysis
The experiments evaluate DeepCervix on annotated cervical-cell images from SIPAKMED. The dataset supports binary, ternary, and five-category classification with fixed training, validation, and test partitions.
- Dataset description: SIPAKMED contains 4049 annotated cervical pap-smear cell images classified by expert cytopathologists into five morphological categories.The categories are superficial-intermediate, parabasal, koilocytotic, metaplastic, and dyskeratotic.
5.2 Data setting
The SIPAKMED experiment uses 4,049 annotated cervical cell images, partitioned by class into training, validation, and test sets and augmented during training.
- 4,049 annotated cervical cell images comprise the SIPAKMED dataset.
- Each class is divided into 60% training, 20% validation, and 20% testing data.
- The study evaluates 2-class, 3-class, and 5-class cervical-cell classification.The five-class categories are superficial, parabasal, koilocytotic, metaplastic, and dyskeratotic; the three-class categories are Normal, abnormal, and benign.
- Data augmentation increases the training dataset by a factor of 6.
5.3 Experimental setup
The models are trained and tested in Google Colaboratory, a Jupyter-based cloud environment configured with machine-learning libraries and an NVIDIA Tesla K80 GPU.
- Google Colaboratory provides the Jupyter-based cloud environment used to train and test the model.
- The environment includes TensorFlow, Matplotlib, Keras, PyTorch, and OpenCV for deep-learning experiments.
- An NVIDIA Tesla K80 GPU supplies the Colab runtime used for deep learning.
5.4 Evaluation method
The study evaluates cervical-cell classification using precision, recall, F1 score, and accuracy to characterize complementary aspects of model performance.
- Precision, recall, F1 score, and accuracy are used as the classification evaluation metrics.
- Precision measures correctly identified samples among recognized representations, while recall measures recognition of all relevant samples.
- The F1 score combines precision and recall through their harmonic mean.
- Accuracy measures the proportion of correctly predicted samples among all samples.
5.5 Results and analysis
DeepCervix is evaluated against individual fine-tuned deep-learning models and late fusion across binary, three-class, and five-class cervical-cell classification. HDFF achieves the highest reported accuracy across these tasks, with performance remaining strong as class count increases.
- 5.5.1 Evaluation results: HDFF achieves 99.85% accuracy for binary cervical-cell classification, matching late fusion and VGG16.
- 5.5.1 Evaluation results: 99.38% accuracy is achieved by HDFF for three-class classification, exceeding late fusion at 98.52%.
- 5.5.1 Evaluation results: 99.14% accuracy is achieved by HDFF for five-class classification, followed by late fusion at 98.64%.
- 5.5.1 Evaluation results: HDFF obtains the highest classification accuracy among the evaluated methods for binary, three-class, and five-class problems.
- 5.5.2 Visualized analysis: For five-class classification, HDFF correctly classifies 805 of 812 images in its confusion matrix.
- 5.5.2 Visualized analysis: HDFF improves five-class accuracy over VGG16, late fusion, VGG19, ResNet50, and XceptionNet by 0.87%, 0.5%, 2.71%, 3.08%, and 33.37%, respectively.
- 5.5.3 Comparative analysis: Compared with existing SIPAKMED studies, the method reports 1.60% higher binary accuracy and 0.19% higher five-class accuracy.
- 5.5.4 Computational analysis: Training each individual deep-learning model takes around six hours for 100 epochs, whereas HDFF training takes a few minutes at 3 seconds per epoch.
5.6 Additional Experiment
The Herlev experiment evaluates DeepCervix on binary and seven-class cervical-cell classification using a fixed train/validation/test split. HDFF achieves the strongest reported accuracy among the compared approaches.
- Dataset and setup: 917 Herlev images are divided into seven classes, which are additionally grouped into benign and malignant categories.The dataset contains 242 benign and 675 malignant images.
- Dataset and setup: 60% of each class is used for training, 20% for validation, and the remainder for testing, with training augmentation increasing the dataset fourteenfold.
- Binary classification: 98.91% accuracy is achieved by HDFF for Herlev binary classification, exceeding LF by 1.08%.
- Seven-class classification: 90.32% accuracy is achieved by HDFF for seven-class Herlev classification, with average precision, recall, and F1 scores of 0.915, 0.911, and 0.916.
- Overall comparison: HDFF obtains the highest classification accuracy across the compared Herlev approaches, following ResNet-50 and LF in performance relative to individual models.
- Overall comparison: Existing seven-class Herlev results range from 68.54% to 95.9%, while the authors report that HDFF outperforms most existing methods.
6 Discussion
The discussion attributes performance differences to dataset and model factors, reports HDFF as the strongest approach, and examines representative misclassifications. Errors occur for visually ambiguous or dark-stained cells.
- Discussion: Imaging modality, image quality, dataset distribution, architecture, complexity, losses, optimization, and training duration influence model performance.
- Discussion: HDFF provides the best SIPAKMED result, while VGG16 performs relatively well among the individual models on that dataset.
- Discussion: 805 of 812 SIPAKMED images are correctly classified in the five-class task, while Koilocytotic and metaplastic cells remain challenging.
- Discussion: ResNet-50 performs better than the other individual models on Herlev, suggesting it is preferable for highly imbalanced and small datasets.
- Error analysis: Misclassifications include dyskeratotic, koilocytotic, metaplastic, carcinoma, columnar, and dysplastic cells with difficult boundaries, similar appearances, or dark staining.
7 Conclusion and Future work
The conclusion presents DeepCervix as an end-to-end HDFF approach evaluated on SIPAKMED and Herlev, with high reported classification accuracies. Future work targets multiclass performance, overlapping cells, additional models, and image noise.
- Conclusion: DeepCervix uses deep feature fusion for end-to-end cervical-cell classification without requiring pre-segmented cytoplasm or nucleus features.
- Conclusion: SIPAKMED accuracies reach 99.85%, 99.38%, and 99.14% for 2-class, 3-class, and 5-class classification, respectively.
- Conclusion: Herlev accuracies reach 98.91% for binary classification and 90.32% for seven-class classification.
- Future work and limitations: Herlev seven-class performance degrades relative to SIPAKMED, and the authors identify this as a limitation of the method.
- Future work and limitations: Future work includes integrating pre-segmented features, evaluating other deep-learning models, generalizing to overlapping cells, and applying denoising methods for noise-related degradation.