A deep learning approach for cervical cell classification and segmentation using CNNs, U-Net, and Detectron2 on the SpikMed dataset.

Cervical cancer remains one of the most prevalent cancers among women worldwide. Early and accurate detection of abnormal cervical cells plays a crucial role in effective diagnosis and treatment.

This project presents a deep learning–based pipeline for cervical cell classification, segmentation, and instance detection, integrating multiple architectures — Convolutional Neural Networks (CNNs) for classification, U-Net for semantic segmentation, and Detectron2 for instance segmentation of overlapping cells. All models were trained and evaluated on the SpikMed dataset, which contains high-resolution microscopic cervical cell images.

The project comprises three core modules designed to work sequentially:

A custom CNN was trained to classify cervical cells into normal and abnormal categories. The network utilized convolutional and pooling layers to extract discriminative features from high-resolution images.

Key Steps:

• Input preprocessing and normalization
• Training with cross-entropy loss
• Evaluation using accuracy and F1-score metrics

To achieve precise cell boundary detection, a U-Net architecture was implemented. The model effectively segmented individual cells from complex background regions, leveraging skip connections to preserve spatial details.

Highlights:

• Encoder–decoder structure based on ResNet backbone
• Data augmentation using Albumentations
• Dice coefficient and IoU metrics for evaluation

For overlapping cell scenarios, Detectron2’s Mask R-CNN was applied. It enabled detection and separation of multiple cell instances within a single image, improving the accuracy of segmentation under real-world conditions.

Configuration:

• Pretrained COCO weights for initialization
• Fine-tuned using custom SpikMed annotations
• Threshold-based filtering for high-confidence masks

Training Details:

• Framework: PyTorch, Detectron2
• Image Size: 256×256
• Optimizer: Adam
• Learning Rate: 1e-4
• Loss Functions: BCEWithLogits (U-Net), CrossEntropy (CNN)
• Epochs: 20–30 depending on convergence

Visual outputs include segmentation masks, class activation maps, and overlapping instance predictions.

cervical-cancer-segmentation/
│
├── notebooks/
│   ├── 1_spikemed_classification.ipynb      # CNN for cell classification
│   ├── 2_unet_segmentation.ipynb            # U-Net training & segmentation
│   ├── 3_detectron_overlap.ipynb            # Detectron2 instance segmentation
│
├── src/
│   ├── unet_model.py                        # U-Net architecture
│   ├── dataset.py                            # Dataset handling
│   ├── utils.py                              # Helper functions (optional)
│
├── data/                                    # Local dataset (not uploaded)
└── outputs/                                 # Predictions, metrics, visualizations

# Clone the repository
git clone https://github.com/yawar-abass/cervical-cancer-seg.git
cd cervical-cancer-segmentation

# Install dependencies
pip install -r requirements.txt

# Open notebooks for training or inference
jupyter notebook notebooks/

• Integrate Transformer-based segmentation architectures (e.g., SegFormer or SwinUNet)
• Deploy model inference via a web-based dashboard for real-time analysis
• Extend dataset for multi-class cervical cell detection
• Explore self-supervised and few-shot learning for limited data scenarios
• Optimize overlapping instance segmentation with improved post-processing techniques