Automatic dual-modality breast tumor segmentation in PET/CT images using CT-guided transformer

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast tumor segmentation is vital for breast cancer diagnosis and treatment planning.
• Current segmentation methods primarily focus on CT images, with limited application to PET/CT.
• Accurate segmentation of PET/CT images is needed to leverage both functional and structural information.

Purpose of the Study:

• Develop a deep learning algorithm for breast tumor segmentation in PET/CT images.
• Integrate functional (PET) and structural (CT) information to enhance segmentation accuracy and speed.
• Assist physicians in patient diagnosis and treatment by providing precise segmentation outcomes.

Main Methods:

• Proposed a CT-Guided Transformer model for automatic breast tumor segmentation in PET images.
• Utilized multi-scale CT features to generate attention maps for PET features.
• Employed similarity-based contrastive learning for effective fusion of multimodal features.

Main Results:

• Achieved superior segmentation performance with 86.19% Dice and 75.73% Jaccard on a clinical dataset.
• Outperformed standard cross-attention methods by 3.86% Dice and 3.64% Jaccard on a public benchmark.
• Demonstrated superior accuracy compared to single-modality and other multimodal fusion techniques.

Conclusions:

• Presented a deep learning method for joint segmentation of PET/CT images.
• Significantly improved breast tumor delineation accuracy over existing methods.
• Showcased the potential of the developed approach for enhanced breast tumor diagnosis.