A multi-task framework for breast cancer segmentation and classification in ultrasound imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Ultrasound (US) is vital for early breast cancer detection.
• Deep learning shows promise for tumor segmentation and classification in US images.
• Current systems face challenges like data standardization issues, exclusion of non-tumor images, and single-task limitations, leading to biased outcomes.

Purpose of the Study:

• To explore the potential of multi-task systems for improving breast cancer lesion detection.
• To develop an end-to-end multi-task framework integrating classification and segmentation.
• To address limitations of existing single-task methodologies in breast cancer US imaging.

Main Methods:

• Introduced an end-to-end multi-task framework to leverage correlations between lesion classification and segmentation.
• Conducted a comprehensive analysis of the BUSI dataset to identify irregularities.
• Developed an algorithm for detecting duplicated images within the dataset to curate it.

Main Results:

• The multi-task framework achieved performance improvements of nearly 15% in both segmentation and classification compared to single-task approaches.
• Experiments utilized a curated dataset to minimize outcome biases.
• A comparative analysis demonstrated statistically significant enhancements over state-of-the-art methods.

Conclusions:

• Multi-task techniques demonstrate superior generalization capabilities across benign, malignant, and non-tumor breast ultrasound images.
• The proposed methodology advances towards more general AI architectures for breast cancer detection.
• The findings suggest real clinical applicability for multi-task systems in breast cancer diagnostics.