Evaluating multi-task network architectures for simultaneous breast lesion segmentation and classification in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast lesion segmentation and classification in ultrasound (US) are crucial for computer-aided breast cancer diagnosis.
• Challenges include lesion variability and poor US image quality, hindering accurate diagnosis.
• Deep learning offers potential but requires optimized approaches for joint tasks.

Purpose of the Study:

• To present and compare multi-task learning (MTL) network configurations for simultaneous breast lesion segmentation and classification in US images.
• To evaluate different feature sharing and integration strategies between segmentation and classification tasks.
• To identify the most effective architectural arrangement for joint breast lesion analysis.

Main Methods:

• Developed and compared multiple MTL network configurations combining SegResNet (segmentation) and EfficientNet (classification).
• Explored various feature sharing and integration schemes between the two tasks.
• Evaluated models on a dataset of 810 2D breast US images from two centers, comparing against state-of-the-art MTL methods.

Main Results:

• The best-performing MTL configuration integrated SegResNet features into an EfficientNet classifier.
• Achieved a Dice coefficient of 81.19% for lesion segmentation and an AUC of 97.27% for classification.
• Demonstrated accurate performance for both segmentation and classification tasks.

Conclusions:

• Combined deep learning architectures using MTL show significant potential for improving breast cancer diagnosis.
• The proposed MTL approach enhances computer-aided diagnosis by accurately segmenting and classifying breast lesions in ultrasound images.
• Optimized feature sharing in MTL networks is key to achieving high performance in joint imaging tasks.