Cervical Cell/Clumps Detection in Cytology Images Using Transfer Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical cancer is a leading cause of cancer-related deaths in women globally.
• Automated screening using deep learning shows promise for improving diagnostic accuracy.
• Limited annotated datasets hinder the training of deep learning models for cervical cell analysis.

Purpose of the Study:

• To address the challenge of limited data in deep learning for cervical cancer screening.
• To propose an optimized transfer learning strategy for cervical cell and clump detection.
• To enhance the performance of deep learning models in detecting cervical abnormalities.

Main Methods:

• Utilized COCO pre-trained models for cervical cell/clump detection tasks, differing from ImageNet pre-training.
• Implemented multi-scale training and analyzed various bounding box loss functions for improved detection.
• Investigated the impact of dataset mean and standard deviation on model performance, optimizing for cervical cell data.

Main Results:

• Achieved a mean Average Precision (mAP) of 61.6% and Average Recall (AR) of 87.7% using a Resnet50 backbone.
• Demonstrated significant performance improvements of 12.8% (mAP) and 23.7% (AR) over existing methods.
• Identified optimal data normalization (mean and std) for cervical cell datasets.

Conclusions:

• The proposed transfer learning approach using COCO pre-trained models effectively addresses data limitations in cervical cancer detection.
• Optimized training strategies, including multi-scale training and appropriate loss functions, enhance model performance.
• This method offers a significant advancement in automated cervical cancer screening, improving diagnostic accuracy and patient outcomes.