CellGAN: Conditional Cervical Cell Synthesis for Augmenting Cytopathological Image Classification

Zhenrong Shen, Maosong Cao, Sheng Wang, Lichi Zhang, Qian Wang

Abstract

Automatic examination of thin-prep cytologic test (TCT) slides can assist pathologists in finding cervical abnormality for accurate and efficient cancer screening. Current solutions mostly need to localize suspicious cells and classify abnormality based on local patches, concerning the fact that whole slide images of TCT are extremely large. It thus requires many annotations of normal and abnormal cervical cells, to supervise the training of the patch-level classifier for promising performance. In this paper, we propose CellGAN to synthesize cytopathological images of various cervical cell types for augmenting patch-level cell classification. Built upon a lightweight backbone, CellGAN is equipped with a non-linear class mapping network to effectively incorporate cell type information into image generation. We also propose the Skip-layer Global Context module to model the complex spatial relationship of the cells, and attain high fidelity of the synthesized images through adversarial learning. Our experiments demonstrate that CellGAN can produce visually plausible TCT cytopathological images for different cell types. We also validate the effectiveness of using CellGAN to greatly augment patch-level cell classification performance.

Introduction

Cervical cancer screening relies heavily on the Thin-prep Cytologic Test (TCT), a process that requires pathologists to identify a handful of abnormal cells among thousands of normal ones within a massive Whole Slide Image (WSI). While deep learning models offer a promising path toward automating this "needle in a haystack" problem, they are fundamentally limited by the availability of high-quality, annotated data.

In clinical practice, abnormal cervical cells—particularly those representing high-grade lesions—are significantly rarer than normal cells. This creates a severe class imbalance that hinders the training of robust patch-level classifiers. Traditional data augmentation techniques, such as rotation or flipping, fail to introduce the morphological diversity necessary for a model to distinguish subtle pathological changes. To bridge this gap, we present CellGAN, a specialized generative adversarial network designed to synthesize realistic, class-specific cervical cell images to augment training datasets.

Methodology

CellGAN is built upon the FastGAN backbone, chosen for its training stability and computational efficiency. The architecture introduces three primary innovations to ensure the synthesized images are both biologically plausible and useful for classification tasks:

  1. Non-linear Class Mapping Network: Rather than directly feeding class labels into the generator, CellGAN utilizes a dedicated network to embed labels into a high-dimensional space. These embeddings modulate image features across multiple scales using Adaptive Instance Normalization (AdaIN), allowing for precise control over cellular attributes specific to each pathological stage.
  2. Skip-layer Global Context (SGC) Module: Cervical cells in TCT slides often appear in complex arrangements—isolated, squeezed, or overlapping. The SGC module captures long-range spatial dependencies, ensuring that the synthesized clusters reflect the realistic spatial relationships found in clinical samples.
  3. Multi-faceted Discriminator: To maintain high fidelity, the discriminator employs a PatchGAN architecture to focus on local texture and color consistency. It also utilizes a projection-based method to match conditional distributions and self-supervised reconstruction decoders as a regularization technique to prevent mode collapse.

Experimental Setup

The model was trained and evaluated using a dataset of 14,477 images (256×256 pixels) sourced from three clinical centers. The images were categorized into five squamous cell types according to the Bethesda System (TBS):

  • NILM: Normal/Negative (7,662 images)
  • ASC-US: Atypical squamous cells of undetermined significance (2,275 images)
  • LSIL: Low-grade squamous intraepithelial lesion (2,480 images)
  • ASC-H: Atypical squamous cells, cannot exclude HSIL (1,638 images)
  • HSIL: High-grade squamous intraepithelial lesion (422 images)

The dataset exhibits a clear long-tail distribution, with nearly 20 times more NILM images than HSIL images, emphasizing the need for synthetic augmentation of the rarer classes.

Results and Discussion

Generative Performance

CellGAN demonstrated superior performance in generating high-fidelity images compared to established benchmarks. It achieved a Fréchet Inception Distance (FID) of 41.51, outperforming both BigGAN (47.32) and Latent Diffusion Models (LDM) (60.65). While LDMs are often considered the state-of-the-art for general image synthesis, they struggled in this domain to maintain clear cell boundaries and class separability given the specialized nature of medical data.

Classification Improvement

The primary value of CellGAN lies in its ability to improve downstream classification. When synthetic images were used to augment the training of standard classifiers, the performance gains were substantial:

  • ResNet-34: Accuracy increased from 74.30% to 79.55% (+5.25%).
  • DenseNet-121: Accuracy increased from 72.10% to 76.15% (+4.05%).

The model proved particularly effective at capturing the distinct morphological features of HSIL cells, which are critical for accurate cancer diagnosis but difficult to collect in large quantities.

Conclusion and Future Work

CellGAN addresses the critical data bottleneck in cervical cancer screening by providing a reliable method for synthesizing rare pathological cell types. By incorporating global context and non-linear class mapping, the framework produces images that are both visually and statistically similar to real clinical data.

Future research will focus on increasing the output resolution beyond 256×256 pixels and developing methods for fine-grained control over specific cellular attributes, such as the nucleus-to-cytoplasm ratio. Additionally, we envision CellGAN serving as an interactive educational tool for training junior pathologists in the identification of rare cervical abnormalities.

Summary updated

Abstract

Automatic examination of thin-prep cytologic test (TCT) slides can assist pathologists in finding cervical abnormality for accurate and efficient cancer screening. Current solutions mostly need to localize suspicious cells and classify abnormality based on local patches, concerning the fact that whole slide images of TCT are extremely large. It thus requires many annotations of normal and abnormal cervical cells, to supervise the training of the patch-level classifier for promising performance. In this paper, we propose CellGAN to synthesize cytopathological images of various cervical cell types for augmenting patch-level cell classification. Built upon a lightweight backbone, CellGAN is equipped with a non-linear class mapping network to effectively incorporate cell type information into image generation. We also propose the Skip-layer Global Context module to model the complex spatial relationship of the cells, and attain high fidelity of the synthesized images through adversarial learning. Our experiments demonstrate that CellGAN can produce visually plausible TCT cytopathological images for different cell types. We also validate the effectiveness of using CellGAN to greatly augment patch-level cell classification performance.
BibTeX
@inproceedings{shen2023cellgan,
  author = {Zhenrong Shen and Maosong Cao and Sheng Wang and Lichi Zhang and Qian Wang},
  title = {{CellGAN: Conditional Cervical Cell Synthesis for Augmenting Cytopathological Image Classification}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  url = {https://openreview.net/forum?id=U7RTNDx0Jg},
}
Figure