Deep Learning for Breast Cancer Diagnosis from Mammograms-A Comparative Study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep convolutional neural networks (CNNs) show promise in medical image analysis.
• Computer-aided diagnosis (CADx) systems aim to improve the accuracy and efficiency of disease detection.
• Breast cancer diagnosis relies heavily on accurate interpretation of mammographic data.

Purpose of the Study:

• To evaluate the performance of state-of-the-art CNNs for breast cancer mass lesion classification.
• To compare the effectiveness of two training strategies: using pre-trained weights versus random initialization.
• To determine the optimal approach for CNN implementation in mammography-based CADx systems.

Main Methods:

• Training and evaluating multiple CNN architectures on two distinct mammographic datasets.
• Utilizing regions of interest (ROIs) containing benign and malignant mass lesions.
• Conducting comparative analysis under two training scenarios: transfer learning (pre-trained weights) and training from scratch (random initialization).

Main Results:

• CNNs initialized with pre-trained weights consistently outperformed those trained from scratch across both datasets.
• Fine-tuning pre-trained networks demonstrated superior performance in distinguishing benign from malignant mass lesions.
• The study confirmed the advantage of leveraging knowledge from pre-existing large-scale datasets.

Conclusions:

• Transfer learning by fine-tuning pre-trained CNNs is a highly effective strategy for breast cancer CADx.
• This approach offers a significant advantage over training deep learning models from random initialization for mammographic analysis.
• The findings support the integration of fine-tuned pre-trained CNNs into clinical workflows for improved breast cancer diagnosis.