Mitosis detection in histopathological images using customized deep learning and hybrid optimization algorithms.

Area of Science:

• Pathology
• Computational Biology
• Machine Learning

Background:

• Accurate mitosis identification is vital for cancer diagnosis.
• Challenges include class imbalance and morphological variations in histopathological images.
• Existing methods struggle with precise mitosis detection.

Purpose of the Study:

• To develop a Customized Deep Learning (CDL) model for enhanced mitosis detection in histopathological images.
• To address challenges of class imbalance and improve localization accuracy.
• To leverage transfer learning and hybrid optimization algorithms for superior performance.

Main Methods:

• A Customized Deep Learning (CDL) model integrating transfer learning and hybrid CNN architecture.
• Utilized skip connections for improved mitosis localization.
• Employed a hybrid optimization mechanism combining Jellyfish Search Optimizer (JSO) and Walrus Optimization Algorithm (WOA).

Main Results:

• The CDL model achieved an F1 score of 0.994 and 98.8% accuracy on benchmark datasets.
• Demonstrated superior performance in detecting mitotic figures.
• Effectively countered class imbalance and improved feature extraction.

Conclusions:

• The CDL model offers a robust and highly effective solution for mitosis detection.
• Significantly aids pathologists in accurate cancer diagnosis and prognosis.
• Shows potential for real-time applications and broader histopathological analyses.