Source-linked AI summary
CVM-Cervix: A Hybrid Cervical Pap-Smear Image Classification Framework Using CNN, Visual Transformer and Multilayer Perceptron
Wanli Liu, Chen Li, Ning Xu, Tao Jiang, Md Mamunur Rahaman, Hongzan Sun, Xiangchen Wu, Weiming Hu, Haoyuan Chen, Changhao Sun, Yudong Yao, Marcin Grzegorzek
TL;DR
Cervical cancer mortality remains very high in underdeveloped countries, motivating computer-aided diagnosis. CVM-Cervix combines CNN, Visual Transformer, and Multilayer Perceptron modules for cervical cell classification, obtaining 91.72% accuracy and compressing model parameters from 120 MB to 60 MB.
Problem
Cervical cancer mortality is very high in underdeveloped countries, motivating the emergence of automatic computer-aided diagnosis systems.
Method
CVM-Cervix uses CNN and Visual Transformer modules for local and global feature extraction, then fuses these features with a Multilayer Perceptron.
Results
91.72% accuracy was obtained on the CRIC and SIPaKMeD datasets, while model parameter size was compressed from 120 MB to 60 MB.
Takeaways & Limitations
CVM-Cervix demonstrates effectiveness and potential for cervical Pap-smear image classification, with a more efficient compressed model in an extended experiment.
Takeaways & Limitations
CVM-Cervix is currently limited to the classification task of cell images.
Abstract
from arXiv · showhide
Cervical cancer is the seventh most common cancer among all the cancers worldwide and the fourth most common cancer among women. Cervical cytopathology image classification is an important method to diagnose cervical cancer. Manual screening of cytopathology images is time-consuming and error-prone. The emergence of the automatic computer-aided diagnosis system solves this problem. This paper proposes a framework called CVM-Cervix based on deep learning to perform cervical cell classification tasks. It can analyze pap slides quickly and accurately. CVM-Cervix first proposes a Convolutional Neural Network module and a Visual Transformer module for local and global feature extraction respectively, then a Multilayer Perceptron module is designed to fuse the local and global features for the final classification. Experimental results show the effectiveness and potential of the proposed CVM-Cervix in the field of cervical Pap smear image classification. In addition, according to the practical needs of clinical work, we perform a lightweight post-processing to compress the model.
1. Introduction
Cervical cytopathology screening is laborious and error-prone, motivating computer-aided diagnosis. CVM-Cervix combines CNN, Visual Transformer, and MLP modules for 11-class cervical-cell classification, achieving 91.72% accuracy with lightweight compression.
- Motivation: Manual inspection is difficult because each slide contains thousands of cells and experts are prone to mistakes.
- Motivation: Computer-aided diagnosis can analyze Pap slides quickly and accurately, addressing the burdens of manual screening.
- Proposed framework: CVM-Cervix combines CNN and Visual Transformer modules to extract local and global features, then uses MLP for feature fusion and classification.
- Task and data: The framework combines SIPaKMeD and CRIC datasets for an 11-category single-cell cervical cytopathology classification task.
- Results and deployment: 91.72% accuracy is reported for CVM-Cervix, and lightweight post-processing compresses the model for clinical needs.
2. Related work
Related work includes diverse CNN, Visual Transformer, attention, and feature-fusion approaches for image analysis and cervical-cell classification. Existing cervical-cell studies report strong results on narrower binary or five-class tasks, while the paper identifies limited CNN–VT integration for cervical classification.
- CNN methods: CNN architectures such as Xception, VGG, ResNet, and DenseNet support image feature extraction, efficiency, deeper networks, residual learning, and direct layer connections.
- Feature fusion and attention: Other image-classification methods use attention, multi-scale global and local streams, or hybrid deep feature fusion to emphasize informative regions and combine representations.
- Visual Transformer methods: Visual Transformer variants apply attention-based mechanisms to vision, with ViT, BoTNet, DeiT, and T2T-ViT targeting accuracy, global information, or reduced resources.
- Cervical-cell classification: For cervical datasets, evaluated ResNet and VGG accuracies varied between 84–86%, with ResNet showing higher accuracy and smaller standard deviation.
- Cervical-cell classification: A prior cervical-cell method achieved 99.85% accuracy on binary SIPaKMeD classification, while a GCN method achieved 98.37% on a five-class task.
- Identified gap: The paper states that existing studies mostly use CNNs and none combines CNN and VT models for cervical-cell classification tasks.
3. Method
CVM-Cervix combines CNN and visual-transformer modules for local and global feature extraction, then fuses their features with an MLP for classification. The pipeline preprocesses cervical-cell images and applies lightweight post-processing for model compression.
- Preprocessing: CVM-Cervix preprocesses images through resizing, random horizontal flipping, and training-data normalization.Images are resized to 224 × 224 pixels, with a 50% probability of horizontal flipping.
- Feature extraction: The framework uses CNN and visual-transformer modules to extract local and global features, respectively.The CNN uses Xception, while the visual-transformer module uses a tiny DeiT model without distillation tokens.
- Model compression: A lightweight post-processing module is applied to compress the trained model.
- CNN module: The CNN module uses depthwise separable convolution, which has lower parameter count and operation cost than conventional convolution.The CNN extracts 2048-dimensional features using the Xception structure without its final classification layer.
- VT module: The visual-transformer module converts images into patch sequences and uses Transformer encoding to extract 192-dimensional features.Positional information and classification headers are added before the sequence enters the Transformer encoder.
4. Experiments and analysis
The experiments combine CRIC and SIPaKMeD cervical-cell datasets into an 11-category dataset. Images are divided into training, validation, and test subsets for evaluation.
- CRIC dataset: CRIC contributes cropped cervical-cell images manually classified into six Bethesda-system categories.The supplied passages identify the categories as ASC-H, ASC-US, SCC, HSIL, LSIL, and NILM.
- SIPaKMeD dataset: SIPaKMeD contains 4049 cropped cervical-cell images across five categories: dyskeratotic, koilocytotic, metaplastic, parabasal, and superficial intermediate.
- Combined dataset: The combined dataset contains six CRIC categories and five SIPaKMeD categories, totaling 11 categories.CRIC contributes six categories, while SIPaKMeD contributes five.
- Combined dataset: The combined dataset contains 8838 images assembled from CRIC and SIPaKMeD.
- Data partitioning: Images are randomly split within each class into 60% training, 20% validation, and 20% test sets.
4.3. Evaluation methods
CVM-Cervix is evaluated using precision, recall, F1-Score, and accuracy. The paper defines these measures using predicted positives, negatives, and classification correctness.
- Evaluation metrics: Precision, recall, F1-Score, and accuracy are the evaluation metrics used for CVM-Cervix.
- Confusion-matrix terms: True Positive and True Negative denote correctly predicted positive and negative samples, while False Positive and False Negative denote the corresponding errors.
- Evaluation metrics: Precision is the proportion of true positives among all positive predictions, whereas recall is the ratio of predicted positive samples to total positive samples.
- Evaluation metrics: F1-Score combines precision and recall, while accuracy is the ratio of correct predictions to the total number of samples.
4.4. Experimental results and analysis
CVM-Cervix achieves broadly similar validation and test performance, with test accuracy of 91.72%, and exceeds existing methods according to the paper. Confusion matrices show strongest classification for several categories and recurring confusions among the second, fifth, sixth, seventh, and eighth categories.
- Performance metrics: 92.80% precision, 92.90% recall, 92.80% F1-Score, and 92.87% accuracy are reported on the validation set.
- Performance metrics: 91.80% precision, 91.60% recall, 91.70% F1-Score, and 91.72% accuracy are reported on the test set.
- Comparison and generalization: The validation-to-test performance difference is about 1%, and the paper states that CVM-Cervix surpasses existing methods on the combined-dataset test set.
- Confusion-matrix analysis: The validation confusion matrix shows very accurate classifications for the first, third, and sixth through eleventh categories.
- Confusion-matrix analysis: The main validation confusions involve the second and fifth categories, while the test matrix shows similar errors involving the second, fifth, sixth, seventh, and eighth categories.Reported test errors include second-to-fifth, second-to-sixth, fifth-to-second, fifth-to-sixth, and seventh-to-eighth misclassifications.
- Abnormal-type classification: The paper identifies the first, third, fourth, seventh, and eighth categories as abnormal and reports higher classification accuracy for abnormal types.The motivation emphasizes avoiding misclassification of cancer patients in practical applications.
4.5. Computational time
CVM-Cervix was trained for 100 epochs on 5,308 training images and evaluated on 1,763 test images. Although training took 2.3 hours, testing required only 31 seconds.
- 100 epochs were used to train CVM-Cervix on 5,308 training images.
- 2.3 hours was the total training time, with approximately 82 seconds required per epoch.
- 31 seconds was required to test CVM-Cervix on 1,763 test images.
4.6. Lightweight post-processing
A lightweight post-processing procedure compresses CVM-Cervix by quantizing weights from 32 to 16 bits. The compressed model is smaller and faster to train, with classification performance nearly unchanged.
- 60 MB was the resulting model parameter size after quantization.
- Around -1% classification-performance change accompanied the compression, while total training time shortened to around 1 hour.
4.7. Extended experiment
Extended experiments evaluated CVM-Cervix with augmentation and on Herlev and peripheral blood cell datasets. The framework achieved strong results across these settings, including 91.72% test accuracy on the combined dataset and 98.89% validation accuracy on peripheral blood cells.
- Combined dataset: 22 comparison models included 18 CNNs and four visual-transformer models under the same experimental settings.
- Combined dataset: 91.72% was CVM-Cervix’s highest test accuracy, nearly 3% above DenseNet169 and DeiT.
- Combined dataset: 91.80% precision, 91.60% recall, and 91.70% F1-Score were reported for CVM-Cervix on the combined dataset.
- Data augmentation: 93.60% precision, 93.50% recall, 93.50% F1-Score, and 93.48% accuracy were obtained after training-set augmentation, roughly 2% above the unaugmented setting.
- Herlev dataset: On Herlev, CVM-Cervix performed slightly less well on the test set because of too little data and uneven class counts, but still achieved good performance.
- Peripheral blood cell dataset: On the peripheral blood cell dataset, validation precision, recall, F1-Score, and accuracy were 99.00%, 98.90%, 98.90%, and 98.89%, respectively.
5. Discussion
The discussion attributes classification errors to visually similar or ambiguous cells and difficult imaging conditions. It also identifies a scope limitation: CVM-Cervix has been evaluated only for cell-image classification.
- Misclassification analysis: t-SNE visualization maps the 2,240-dimensional feature vector into three dimensions to inspect extracted features and misclassified images.
- Misclassification analysis: ASC-US images were misclassified as LSIL and NILM, while LSIL images were classified as ASC-US or NILM.
- Misclassification analysis: Light-colored cells, unclear boundaries, dark backgrounds, and blurred nucleus–cytoplasm boundaries made feature extraction inaccurate.
- Misclassification analysis: CVM-Cervix cannot effectively identify cell nuclei in some image-feature extraction cases.
- Misclassification analysis: Similarity and ambiguity between cell images were identified as the main reasons for classification errors.
- Limitations: CVM-Cervix has currently been applied only to cell-image classification and has not been tested on other tasks.
- Implementation choices: The framework uses a tiny DeiT module without distillation tokens, AdamW, ReLU, and cross-entropy because these choices fit the proposed architecture’s efficiency and computational-cost requirements.
6. Conclusion and future work
CVM-Cervix combines CNN and visual-transformer modules for local and global feature extraction, fusing their features with an MLP for cervical cell classification. It achieves 91.72% accuracy on an 11-class combined dataset, generalizes across other image datasets, and supports lightweight model compression.
- Conclusion: CVM-Cervix combines CNN and visual-transformer modules for local and global feature extraction, respectively.
- Conclusion: An MLP fuses the local and global features for final cervical cell classification.
- Experimental results: 91.72% accuracy is obtained for the 11-class classification problem on a combined CRIC and SIPaKMeD dataset.
- Model compression: Model parameter size is compressed from 120 MB to 60 MB through lightweight post-processing for clinical needs.
- Experimental results: CVM-Cervix performs well on other cervical Pap-smear and cell-image datasets, indicating strong generalization ability.
- Future work: Future work includes additional model combinations, module-structure modifications, preprocessing techniques, and comparisons with capsule networks.