Distillation of multi-class cervical lesion cell detection via synthesis-aided pre-training and patch-level feature...

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Healthcare

Background:

• Automated detection of abnormal cervical cells from Thin-Prep Cytologic Test (TCT) images is vital for efficient computer-aided diagnosis (CADx) systems in cervical cancer screening.
• Current CADx model development faces challenges including class imbalance in training data and incomplete annotations.
• Existing methods often neglect crucial visual feature correlations among cells, a key factor pathologists use for accurate identification.

Purpose of the Study:

• To propose a novel distillation framework for enhancing automated cervical cell detection in TCT images.
• To address limitations of class imbalance, incomplete annotations, and overlooked cell feature correlations in existing detection models.
• To develop a versatile framework applicable to various image-level detection networks without altering their inference architecture.

Main Methods:

• A Balanced Pre-training Model (BPM) was developed for patch-level cervical cell classification, utilizing an image synthesis model to create a class-balanced dataset.
• Score Correction Loss (SCL) was designed to facilitate knowledge distillation from the BPM to the detection network, mitigating issues from incomplete annotations.
• Patch Correlation Consistency (PCC) strategy was introduced to leverage correlations among extracted cells, simulating cytopathologists' analytical approach.

Main Results:

• The proposed distillation framework demonstrated superior performance in automated cervical cell detection.
• The method proved adaptable and effective across various detection architectures.
• Experiments on both public and private datasets validated the efficacy of the BPM, SCL, and PCC strategies.

Conclusions:

• The developed distillation framework effectively overcomes challenges in training data preparation for cervical cell detection.
• The integration of class-balanced pre-training, score correction, and patch correlation consistency significantly improves detection accuracy.
• This approach offers a robust and adaptable solution for advancing computer-aided diagnosis in cervical cancer screening.