A novel multilevel iterative training strategy for the ResNet50 based mitotic cell classifier.

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning in oncology

Background:

• Accurate mitotic cell counting is vital for invasive breast cancer grading.
• Manual identification by pathologists is challenging and time-consuming.
• Deep learning models offer automated solutions but face training challenges like local optima.

Purpose of the Study:

• To introduce a novel multilevel iterative training strategy for deep learning models.
• To address the problem of converging to local optimal solutions during model training.
• To improve the accuracy of automatic mitotic cell identification in histopathological images.

Main Methods:

• Development of a multilevel iterative training strategy.
• Construction of a mitotic cell classification model using ResNet50 architecture.
• Comparison of the proposed training strategy against conventional methods on an independent test set.

Main Results:

• Models trained with the proposed strategy demonstrated superior performance compared to conventional strategies.
• The ResNet50 model, trained with the novel strategy and Adam optimizer, achieved an 89.26% F1 score on the MITOSI14 dataset.
• This F1 score surpasses existing state-of-the-art methods reported in the literature.

Conclusions:

• The proposed multilevel iterative training strategy is effective in enhancing deep learning model performance for mitotic cell detection.
• This approach helps overcome local optima issues in model training.
• The method shows significant potential for improving automated breast cancer grading.