BIRADS features-oriented semi-supervised deep learning for breast ultrasound computer-aided diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Computer-Aided Diagnosis

Background:

• Accurate diagnosis of breast ultrasound (US) images is crucial for early cancer detection.
• Small training datasets pose a significant challenge for developing effective deep learning models in medical imaging.
• Integrating clinical knowledge, such as Breast Imaging Reporting and Data System (BIRADS) features, can enhance diagnostic accuracy.

Purpose of the Study:

• To develop a novel BIRADS-SSDL network for accurate breast US image diagnosis with limited training data.
• To integrate clinically-approved BIRADS features into a task-oriented semi-supervised deep learning framework.
• To evaluate the performance and generalizability of the proposed network compared to existing methods.

Main Methods:

• Breast US images were converted into BIRADS-oriented feature maps (BFMs) using distance transformation and Gaussian filtering.
• A semi-supervised deep learning (SSDL) network, incorporating unsupervised stacked convolutional auto-encoder (SCAE) image reconstruction guided by lesion classification, was developed.
• The BIRADS-SSDL network was trained using an alternative learning strategy balancing reconstruction and classification errors.

Main Results:

• The BIRADS-SSDL network achieved superior classification accuracy (around 94.23% and 84.38% on two datasets) compared to conventional SCAE and SSDL methods.
• The developed BIRADS-SSDL demonstrated generalizability across different institutions and US devices without overfitting.
• Classification accuracy was sensitive to segmentation accuracy, highlighting the importance of precise lesion boundary detection.

Conclusions:

• The proposed BIRADS-SSDL network is a promising approach for effective breast US computer-aided diagnosis, particularly with small datasets.
• Integrating BIRADS features enhances the performance of deep learning models for breast lesion classification.
• The BIRADS-SSDL network exhibits robustness and generalizability, making it suitable for diverse clinical settings.