Transfer Learning Fine-Tuning Strategies for Enhanced Ultrasound Breast Cancer Classification

Artificial intelligence (AI) has emerged as a powerful tool in medical image analysis, revolutionizing the field of radiology and improving diagnostic accuracy and efficiency. Among various imaging modalities, ultrasound imaging is crucial in diagnosing a wide range of medical conditions due to its non-invasive nature and real-time imaging capabilities. However, the scarcity of labeled training data and the challenge of constructing effective learning frameworks pose significant hurdles in developing accurate and robust AI models for ultrasound image analysis. This research paper presents a study conducted on ultrasound images, specifically for breast cancer classification, and focuses on the application of Transfer Learning (TL) using state-of-the-art ImageNet pre-trained models including VGG16, VGG19, ResNet50, ResNet101, and InceptionV3. The study also explores the impact of different fine-tuning strategies on the final classification outcome. Strategies such as freezing 100% of layers, freezing 50% of layers, and scratch fine-tuning by training all layers were implemented along with the same common neural network-based classifier built on top. For reproducibility, publicly accessible datasets were used, namely Mendeley Breast and BUSI datasets. Additionally, a stratified 5-fold cross-validation technique was implemented to evaluate the pre-trained models, and metrics such as Accuracy, Sensitivity, Specificity, Precision, and False Positive Rate (FPR) were computed accordingly. This paper demonstrates the necessity of choosing the appropriate fine-tuning strategy aligned with the pre-trained model used. This can eventually enhance the feature extraction task, thus saving time and effort when implementing such an automatic classification framework for ultrasound images. In the application of breast ultrasound cancer classification, InceptionV3 has been found to be the consistent model across all strategies. Fine-tuning 50% of layers for this model has proved to have better performance.