Multimodal Spatial Attention Module for Targeting Multimodal PET-CT Lung Tumor Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiomics

Background:

• Multimodal positron emission tomography-computed tomography (PET-CT) is crucial for cancer assessment, combining PET's sensitivity and CT's anatomical detail.
• Automated tumor segmentation in PET-CT is challenging, with current methods exhibiting low performance, leading to manual segmentation that is time-consuming and inconsistent.
• Existing automated approaches often fail to fully leverage PET's tumor-detecting sensitivity to guide segmentation.

Purpose of the Study:

• To develop and validate a novel deep learning framework for enhanced multimodal PET-CT tumor segmentation.
• To introduce a multimodal spatial attention module (MSAM) designed to automatically focus on tumor regions within PET data.
• To improve the accuracy and efficiency of automated tumor segmentation in clinical cancer imaging.

Main Methods:

• A deep learning framework incorporating a multimodal spatial attention module (MSAM) was developed for PET-CT segmentation.
• The MSAM learns to highlight tumor-indicative regions in PET and suppress physiological uptake in normal tissues.
• Attention maps generated by MSAM guide a convolutional neural network (CNN) backbone for segmenting tumor likelihood in CT images.

Main Results:

• The proposed framework demonstrated effectiveness across two distinct cancer types: non-small cell lung cancer (NSCLC) and soft tissue sarcoma (STS).
• Experimental results showed the MSAM, when integrated with a U-Net backbone, significantly improved segmentation performance.
• The approach surpassed the state-of-the-art lung tumor segmentation method by 7.6% in Dice similarity coefficient (DSC).

Conclusions:

• The developed deep learning framework with MSAM offers a significant advancement in automated PET-CT tumor segmentation.
• The MSAM effectively utilizes PET information to guide accurate segmentation, addressing limitations of previous methods.
• This approach holds promise for improving diagnostic accuracy and efficiency in cancer assessment using PET-CT imaging.