BCDnet: Parallel heterogeneous eight-class classification model of breast pathology.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading cause of mortality in women, necessitating accurate diagnostic tools.
• Current pathological tissue image classification often focuses on binary (benign/malignant) distinctions, which is insufficient for comprehensive diagnosis.
• Small and imbalanced datasets pose challenges for developing robust deep learning models in breast pathology.

Purpose of the Study:

• To develop an advanced deep learning model for the eight-class classification of breast pathological tissue images.
• To improve the accuracy and efficiency of breast cancer diagnosis using computer vision.
• To address the limitations of existing classification methods and small medical datasets.

Main Methods:

• Designed BCDnet, an eight-class classification model integrating deep convolutional neural networks (VGG16 and Resnet50), model ensembling, transfer learning, and feature fusion.
• Employed image segmentation for data augmentation and non-repeated random cropping for dataset balancing.
• Conducted comparative experiments using balanced and unbalanced datasets against pre-trained Resnet50 and VGG16 models.

Main Results:

• The BCDnet model achieved an outstanding correct recognition rate exceeding 98% for eight-class classification.
• Demonstrated superior performance compared to Resnet50 and VGG16 fine-tuning approaches, particularly on the challenging BreaKHis dataset.
• Data augmentation and balancing techniques proved effective in enhancing model performance.

Conclusions:

• The proposed BCDnet model offers a highly accurate and effective solution for multi-class breast pathology image classification.
• The data improvement strategies are crucial for building reliable AI models with limited and imbalanced medical datasets.
• This research provides a valuable tool to assist pathologists in rapid and precise breast cancer diagnosis.