DeepPap: Deep Convolutional Networks for Cervical Cell Classification.

Area of Science:

• Medical imaging
• Computational pathology
• Oncology

Background:

• Cervical cancer screening relies on cell imaging, but accurate segmentation is challenging.
• Traditional methods use hand-crafted features, limiting classification accuracy.
• Existing automation-assisted tools face segmentation difficulties with cell clusters and pathologies.

Purpose of the Study:

• To develop a novel method for direct cervical cell classification without prior segmentation.
• To leverage deep features from convolutional neural networks (ConvNets) for improved accuracy.
• To enhance automation-assisted cervical screening systems.

Main Methods:

• A convolutional neural network (ConvNet) was pretrained on natural images and fine-tuned on cervical cell image patches.
• The method directly classifies cells using deep features, avoiding segmentation steps.
• Image patches were adaptively resampled and centered on nuclei; prediction scores were aggregated during testing.

Main Results:

• The proposed method achieved 98.3% classification accuracy and 0.99 area under the curve on the Herlev benchmark Pap smear dataset.
• Specificity reached 98.3%, outperforming previous algorithms.
• Similar high performance was observed on the H&E stained manual liquid-based cytology (HEMLBC) dataset.

Conclusions:

• Directly classifying cervical cells using deep features with ConvNets overcomes segmentation limitations.
• The method demonstrates superior performance in accuracy, AUC, and specificity for cervical cell classification.
• This approach shows significant promise for developing advanced automation-assisted cervical screening systems.