Explainable artificial intelligence with pyramid vision transformer model for multi-class malignant cell...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Deep learning for medical imaging

Background:

• Digital pathology enhances cancer detection but manual slide analysis is time-consuming and inconsistent.
• Pathologist variability (inter- and intra-operator) affects diagnostic accuracy.
• Deep learning (DL) offers potential for automated analysis in digital pathology.

Purpose of the Study:

• To develop and validate a Hybrid Temporal Deep Learning Network for Automated Malignant Cell Classification in Cytology Slides (HTDL-MCCCS).
• To improve diagnostic consistency, reduce pathologist workload, and enable scalable screening.
• To enhance clinical interpretability and trust in automated diagnostic systems.

Main Methods:

• The HTDL-MCCCS model integrates slide pre-processing, Pyramid Vision Transformer (PVT) for feature learning, and Bidirectional Temporal Recurrent Unit (BiTRU) for classification.
• Pre-processing includes stain normalization, artifact removal, patch extraction, and nuclei segmentation.
• Grad-CAM is used for model interpretability.

Main Results:

• The HTDL-MCCCS model demonstrated superior performance with accuracy rates of 98.16% on the SIPaKMeD dataset and 96.42% on the Herlev dataset.
• The model effectively learns cell-level features and classifies malignant versus benign cells.
• Grad-CAM analysis provided insights into the model's decision-making process.

Conclusions:

• The HTDL-MCCCS model shows significant potential for accurate and consistent automated malignant cell classification in digital pathology.
• This approach can support pathologists, reduce diagnostic variability, and improve efficiency in resource-limited settings.
• The study highlights the utility of hybrid deep learning models in advancing computational pathology.