Breast mass classification in sonography with transfer learning using a deep convolutional neural network and color...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate breast mass classification in ultrasound is crucial for early cancer detection.
• Current diagnostic methods rely on radiologist interpretation, which can be subjective.
• Deep learning offers potential for objective and accurate image analysis.

Purpose of the Study:

• To develop and evaluate a deep learning-based approach for breast mass classification in sonography.
• To compare the deep learning model's performance against experienced radiologists.
• To introduce and assess a novel 'matching layer' for image preprocessing.

Main Methods:

• Employed transfer learning techniques on 882 breast ultrasound images.
• Introduced a 'matching layer' to convert grayscale images to RGB for enhanced feature extraction.
• Utilized back-propagation for fine-tuning and assessed performance with and without color conversion.

Main Results:

• Color conversion significantly improved the area under the receiver operating characteristic curve (AUC) for all transfer learning methods.
• The best-performing model achieved an AUC of 0.936, surpassing radiologists' AUCs (0.806-0.882).
• The approach demonstrated strong performance on two public datasets, achieving AUCs around 0.890.

Conclusions:

• The proposed 'matching layer' is a generalizable technique to enhance deep convolutional neural network performance.
• The deep learning approach shows significant potential to aid radiologists in breast mass classification.
• Further development could lead to a valuable clinical tool for improving diagnostic accuracy in breast sonography.