Artificial classification of cervical squamous lesions in ThinPrep cytologic tests using a deep convolutional neural...

Area of Science:

• Digital Pathology
• Cytopathology
• Artificial Intelligence in Medicine

Background:

• Accurate classification of cervical squamous lesions in ThinPrep cytologic tests (TCT) is crucial for diagnosing squamous cell carcinoma.
• Pathologist interpretation under a microscope is the current standard but can be time-consuming.
• Deep convolutional neural networks (DCNNs) show promise in digital pathology but require validation on diverse datasets.

Purpose of the Study:

• To develop and validate a DCNN model for automated classification of normal and abnormal cervical squamous cells from a multi-center TCT dataset.
• To evaluate the performance of the DCNN model, including its accuracy and efficiency, compared to expert pathologists.
• To assess the impact of an ensemble training strategy (ETS) on the DCNN model's classification capabilities.

Main Methods:

• A VGG16-based DCNN model was trained on a multi-center dataset of 82 TCT samples.
• An ensemble training strategy (ETS) using 5-fold cross-validation was applied.
• Model classifications were compared against diagnoses from two experienced pathologists using paired sample t-tests.

Main Results:

• The DCNN model achieved classification accuracy comparable to that of expert pathologists.
• The ensemble training strategy (ETS) showed a slight, non-significant improvement in accuracy.
• The DCNN model demonstrated a 6-fold increase in speed compared to manual pathologist review.

Conclusions:

• Automated DCNN classification of cervical squamous lesions offers similar accuracy to pathologists but with significantly higher efficiency.
• This automated approach can facilitate wider and more efficient cervical cancer screening.
• Further research is needed to explore the practical implementation of DCNN models in laboratory settings.