Deep Learning for Semantic Segmentation vs. Classification in Computational Pathology: Application to Mitosis...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Current computer-aided diagnostic tools lack efficiency and effectiveness in routine pathology practice.
• Accurate mitosis detection and classification are crucial for breast cancer diagnosis and prognosis.

Purpose of the Study:

• To develop and evaluate two deep learning architectures for efficient and effective mitosis detection and classification in histopathological images.
• To improve upon existing computational methods for breast cancer diagnosis.

Main Methods:

• A two-part method involving image preprocessing and a Convolutional Neural Network (CNN) with handcrafted features for binary classification.
• An end-to-end deep learning methodology utilizing semantic segmentation for mitosis detection and classification.

Main Results:

• The first method achieved 95% accuracy and a 94.35% F1-score, outperforming classical image processing and hybrid CNN approaches.
• The semantic segmentation approach demonstrated over 95% accuracy and a 0.6 Dice index, surpassing existing CNN results (0.9 F1-score).
• Deep learning frameworks are viable for both mitosis detection and classification, showing potential for Whole Slide Image (WSI) analysis.

Conclusions:

• Deep learning architectures offer a promising avenue for developing effective computer-aided systems in pathology.
• The proposed methods demonstrate significant potential for integration into pathologists' daily workflows for breast cancer diagnosis.
• Further research can extend these findings to Whole Slide Images (WSI) and inform the development of advanced computer-aided systems.