Fine tuning deep learning models for breast tumor classification

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate differentiation between benign and malignant breast tumors (BT) is crucial for effective treatment.
• Histopathology images are vital for BT diagnosis, but manual analysis can be subjective and time-consuming.
• Developing automated systems for BT classification can improve diagnostic accuracy and efficiency.

Purpose of the Study:

• To propose and evaluate an approach for enhancing the classification of benign and malignant breast tumors using histopathology images.
• To compare the performance of a custom Convolutional Neural Network (CNN) against pre-trained models for breast tumor differentiation.
• To investigate the impact of metaheuristic optimization algorithms on improving CNN model performance for breast tumor classification.

Main Methods:

• The study utilized the BreakHis dataset comprising histopathology images of breast tumors.
• Preprocessing involved image resizing, data partitioning, and augmentation.
• A custom CNN was developed for feature extraction and classification, with performance compared against pre-trained models (MobileNetV3, EfficientNetB0, Vgg16, ResNet50V2).
• Hyperparameter tuning was performed using Grey Wolf Optimization (GWO) and Modified Gorilla Troops Optimization (MGTO) metaheuristic algorithms.

Main Results:

• The custom CNN model achieved an initial accuracy of 84%, outperforming pre-trained models (74-82%).
• After hyperparameter tuning with MGTO, the custom CNN model reached a significantly improved accuracy of 93.13% within 10 iterations.
• MGTO optimization demonstrated superior performance in enhancing the classification accuracy compared to GWO and unoptimized models.

Conclusions:

• The proposed approach, particularly the custom CNN optimized with MGTO, shows significant potential for accurate and efficient breast tumor classification.
• Automated systems leveraging optimized deep learning models can aid pathologists in distinguishing benign from malignant breast tumors.
• This research contributes to advancing AI-driven diagnostic tools in breast cancer research and clinical practice.