A Novel Multistage Transfer Learning for Ultrasound Breast Cancer Image Classification.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Oncology

Background:

• Artificial intelligence (AI) enhances breast cancer diagnosis, but large medical image datasets are scarce.
• Transfer learning (TL) typically uses ImageNet models, which lack medical image data, limiting its effectiveness.
• Microscopic cancer cell line images are abundant and can supplement limited medical datasets.

Purpose of the Study:

• To investigate if combining natural and medical datasets improves ultrasound breast cancer image classification using AI.
• To develop and evaluate a novel multistage transfer learning (MSTL) algorithm for enhanced breast cancer detection.

Main Methods:

• Implemented a multistage transfer learning (MSTL) algorithm using EfficientNetB2, InceptionV3, and ResNet50 pre-trained models.
• Employed Adam, Adagrad, and stochastic gradient descent (SGD) optimizers.
• Utilized datasets including 20,400 cancer cell images and 600 ultrasound breast cancer images from two sources.

Main Results:

• ResNet50-Adagrad-based MSTL achieved high accuracy: 99 ± 0.612% on the Mendeley dataset and 98.7 ± 1.1% on the MT-Small-Dataset.
• MSTL demonstrated a significant improvement (p=0.01191) over ImageNet-based TL on the Mendeley dataset.
• The developed AI method outperformed existing state-of-the-art techniques for ultrasound breast cancer classification.

Conclusions:

• Combining natural and medical image datasets via MSTL significantly boosts AI performance in ultrasound breast cancer classification.
• This advanced AI approach holds promise for improving early breast cancer diagnosis, particularly in young women.
• The study highlights the potential of leveraging abundant cell line data to overcome medical imaging dataset limitations in AI model training.