Hybrid Loss-Constrained Lightweight Convolutional Neural Networks for Cervical Cell Classification

Wen Chen, Weiming Shen, Liang Gao, Xinyu Li

TL;DR

This paper introduces a novel hybrid loss function combined with lightweight Convolutional Neural Networks (CNNs) to improve the accuracy and efficiency of cervical cell classification. The goal is to make AI-powered cervical cancer screening more accessible and affordable, especially in resource-limited settings, by achieving high performance with significantly reduced computational demands.

What Problem Does This Paper Address?

Cervical cancer is a major global health concern, particularly in developing countries where access to screening and qualified medical professionals is limited. While AI-powered automated screening systems offer a promising solution to increase efficiency and accessibility, their widespread adoption is hampered by high computational costs and resource requirements. Existing deep learning models, while accurate, often demand substantial memory, processing power, and parameters, making them impractical for deployment in low-resource environments. This paper aims to bridge this gap by developing a cost-effective and efficient AI system for cervical cell classification that maintains high accuracy.

Methods

The core of this work lies in a hybrid loss function designed to enhance the discriminative power of CNNs, particularly for distinguishing between visually similar cervical cell types. This hybrid loss combines:

  • Cross-entropy loss: A standard loss function that encourages the network to correctly classify samples.
  • Triplet loss: A loss function that aims to pull features from the same class closer together (intra-class compactness) while pushing features from different classes further apart (inter-class separability). This is crucial for fine-grained classification where subtle differences matter.

This hybrid loss is integrated with lightweight CNN architectures (Squeezenet, MobilenetV2, ShufflenetV2, and Ghostnet). These architectures are chosen for their inherent efficiency in terms of parameters and computational operations. Additionally, label smoothing is employed during training to prevent overfitting. The CNN features from the last convolutional layer are used for triplet loss calculation, while class prediction logits are used for cross-entropy loss.

Datasets

The study utilizes the SIPaKMeD dataset for evaluation. This dataset contains 4049 image samples of isolated cervical cells, categorized into five classes: superficial-intermediate, parabasal, metaplastic, koilocytotic, and dyskeratotic. The images were acquired using a CCD camera and an optical microscope. The paper does not introduce a new dataset but uses this existing benchmark for its experiments.

Key Insights and Ideas

  • Hybrid Loss for Fine-Grained Classification: The combination of cross-entropy and triplet loss is particularly effective for distinguishing between subtle variations in cervical cell morphology, a common challenge in this domain.
  • Lightweight CNNs are Viable: The research demonstrates that efficient, lightweight CNN architectures can achieve performance comparable to larger, more resource-intensive models when paired with an appropriate loss function.
  • Computational Cost is a Major Bottleneck: The paper strongly emphasizes that computational cost is a significant barrier to the real-world deployment of AI in medical imaging, driving the need for efficient models.
  • Triplet Loss Enhances Discriminative Features: The triplet loss component actively works to create a more separable feature space, making it easier for the network to differentiate between confusable cell types.

Key Results

The proposed hybrid loss function, when applied to lightweight CNNs, yielded significant improvements:

  • ShufflenetV2 with the hybrid loss achieved an accuracy of 96.18%, precision of 96.30%, recall of 96.23%, and specificity of 99.08%. Crucially, this was achieved with only one-seventh of the memory usage, one-sixth of the parameters, and one-fiftieth of the total flops compared to DenseNet-121.
  • Ghostnet with the hybrid loss achieved an accuracy of 96.39%, precision of 96.42%, recall of 96.39%, and specificity of 99.09%. This model used half the memory usage, one-quarter of the parameters, and one-fiftieth of the total flops compared to DenseNet-121.
  • While MobilenetV2 did not show the same level of improvement with the hybrid loss, Squeezenet, ShufflenetV2, and Ghostnet all demonstrated enhanced performance with the proposed method compared to traditional cross-entropy loss.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Not Fully End-to-End: The current system focuses on cell classification and is not a complete end-to-end cervical cell detection and screening system. Detecting isolated cells is a separate challenge.
  • Data Scarcity: The reliance on large labeled datasets for deep learning remains a hurdle in healthcare. Unsupervised learning is suggested as a promising avenue to augment data efficiency.
  • Hyperparameter Tuning: The study acknowledges that hyperparameter selection was not exhaustively explored due to limited computational resources. Future work could involve automatic hyperparameter adjustment or machine learning-inspired algorithms for optimization.
  • Broader Applicability: The authors believe the proposed approach can be extended to other medical image processing applications and advocate for further research to establish robust, automated cervical cancer screening systems.

Summary updated

Abstract

Artificial intelligence (AI) technologies have resulted in remarkable achievements and conferred massive benefits to computer-aided systems in medical imaging. However, the worldwide usage of AI-based automation-assisted cervical cancer screening systems is hindered by computational cost and resource limitations. Thus, a highly economical and efficient model with enhanced classification ability is much more desirable. This paper proposes a hybrid loss function with label smoothing to improve the distinguishing power of lightweight convolutional neural networks (CNNs) for cervical cell classification. The results strengthen our confidence in hybrid loss-constrained lightweight CNNs, which can achieve satisfactory accuracy with much lower computational cost for the SIPakMeD dataset. In particular, ShufflenetV2 obtained a comparable classification result (96.18% in accuracy, 96.30% in precision, 96.23% in recall, and 99.08% in specificity) with only one-seventh of the memory usage, one-sixth of the number of parameters, and one-fiftieth of total flops compared with Densenet-121 (96.79% in accuracy). GhostNet achieved an improved classification result (96.39% accuracy, 96.42% precision, 96.39% recall, and 99.09% specificity) with one-half of the memory usage, one-quarter of the number of parameters, and one-fiftieth of total flops compared with Densenet-121 (96.79% in accuracy). The proposed lightweight CNNs are likely to lead to an easily-applicable and cost-efficient automation-assisted system for cervical cancer diagnosis and prevention.
BibTeX
@inproceedings{chen2022hybrid,
  author = {Wen Chen and Weiming Shen and Liang Gao and Xinyu Li},
  title = {{Hybrid Loss-Constrained Lightweight Convolutional Neural Networks for Cervical Cell Classification}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=nxXN0yQ9Olj},
}
Figure