ISA-Net: Improved spatial attention network for PET-CT tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiomics

Background:

• Manual tumor segmentation is time-consuming, costly, and prone to errors.
• Expert-dependent manual annotation leads to significant inter- and intra-observer variability.
• Automated tumor segmentation is crucial for clinical practice and radiomics research.

Purpose of the Study:

• To develop an automated tumor segmentation method using deep learning on multimodal PET-CT data.
• To improve the accuracy and efficiency of tumor segmentation compared to manual methods.
• To leverage the complementary information from PET and CT imaging.

Main Methods:

• Proposed an improved spatial attention network (ISA-Net) for multimodal PET-CT tumor segmentation.
• Utilized multi-scale convolution for feature extraction and attention mechanisms to highlight tumor regions.
• Employed dual-channel inputs in the encoder and fusion in the decoder to integrate PET and CT data.

Main Results:

• ISA-Net demonstrated superior segmentation performance on soft tissue sarcoma (STS) and head and neck tumor (HECKTOR) datasets.
• Achieved Dice Similarity Coefficient (DSC) scores of 0.8378 on STS and 0.8076 on HECKTOR.
• The method showed better generalization capabilities compared to other attention-based segmentation methods.

Conclusions:

• The proposed ISA-Net effectively utilizes multimodal medical image data for accurate tumor segmentation.
• The method capitalizes on the complementary information present in different imaging modalities.
• ISA-Net is adaptable for application to other multimodal or single-modal medical image segmentation tasks.