Saliency-driven system models for cell analysis with deep learning.

Area of Science:

• Computational vision
• Medical image analysis
• Cytopathology

Background:

• Saliency models predict visual attention but their performance with cytopathologists on Pap smear slides is understudied.
• Accurate saliency prediction is crucial for automated detection of cells in challenging conditions like noise and occlusions.

Purpose of the Study:

• To evaluate computational saliency models for predicting cytopathologist eye fixations on Pap smear slides.
• To develop and compare algorithms for retrieving clinically relevant regions of interest (ROIs).
• To support the design of computer-aided diagnosis systems for cytopathology.

Main Methods:

• Recorded eye fixation maps from cytopathologists during routine examinations.
• Compared 13 saliency prediction algorithms, including deep learning models.
• Developed cell-specific convolutional neural networks (CNNs) to analyze bottom-up and top-down saliency factors.

Main Results:

• The proposed cell-specific CNN model demonstrated superior performance over existing methods, significantly reducing false positives.
• The algorithm achieved over 98% accuracy in detecting clinically relevant cells for most diseases, with 87% for carcinoma.
• Bottom-up saliency methods provided satisfactory ROI detection rates (75-86%) for various pathologies.

Conclusions:

• Saliency prediction methods effectively extract ROIs, enabling data reduction for automated Pap smear slide analysis.
• Top-down factors enhance saliency map accuracy, while bottom-up algorithms are useful for predicting cytopathologist fixations.
• This study offers a comparison of saliency models and a method linking conspicuous regions to clinically relevant cells.