Improving Breast Tumor Segmentation in PET via Attentive Transformation Based Normalization

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Positron Emission Tomography (PET) is crucial for cancer diagnosis and treatment monitoring.
• Accurate tumor segmentation is essential for PET-based clinical applications.
• Deep convolutional neural networks (CNNs) are state-of-the-art for PET tumor segmentation, with normalization being a key performance factor.

Purpose of the Study:

• To introduce a new normalization method for improving PET tumor segmentation.
• To address limitations of existing normalization techniques that introduce batch or background noise.
• To enhance the accuracy and efficiency of deep learning models in PET image analysis.

Main Methods:

• Proposed an attentive transformation (AT)-based normalization method.
• Dynamically generated pixel-dependent normalization parameters using channel-wise and spatial-wise attentive responses.
• Exploited breast tumor characteristics in PET images for targeted feature recalibration.

Main Results:

• The AT-based normalization method demonstrated improved breast tumor segmentation performance.
• Effectively recalibrated features pixel-by-pixel, focusing on high-uptake tumor areas.
• Successfully attenuated background noise in PET images, leading to better segmentation outcomes.

Conclusions:

• The proposed AT-based normalization method offers a significant advancement in PET tumor segmentation.
• This technique enhances the performance of deep learning models by intelligently handling image noise.
• The method shows promise for improving clinical applications relying on accurate PET tumor segmentation.