Coco-Attention for Tumor Segmentation in Weakly Paired Multimodal MRI Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Multimodal MRI offers complementary data for lung cancer diagnosis and treatment evaluation.
• Accurate tumor segmentation in chest MRI is challenging due to difficulties in exploiting multimodal information and lack of rigorous registration.

Purpose of the Study:

• To propose a novel coco-attention mechanism for effective exploitation of complementary information in weakly paired multimodal MRI images for accurate lung tumor segmentation.
• To improve tumor localization and segmentation precision in non-small cell lung cancer (NSCLC) using advanced deep learning techniques.

Main Methods:

• A novel coco-attention module comprising multi-modal co-attention (MultiCo-attn) and multi-level coordinate attention (MultiCord-attn) was developed.
• MultiCo-attn uses a bidirectional algorithm to extract complementary features and generate tumor-focused attention maps.
• MultiCord-attn refines segmentation by adjusting feature point weights using multi-level information.

Main Results:

• The proposed coco-attention method demonstrated effectiveness in segmenting lung tumors from weakly paired multimodal MRI scans.
• Significant improvement (p < 0.005) in segmentation accuracy was achieved compared to existing multimodal segmentation methods.
• Ablation experiments confirmed the effectiveness and interpretability of the coco-attention module.

Conclusions:

• The novel coco-attention mechanism provides an effective approach for lung tumor segmentation in multimodal MRI, particularly in scenarios with weakly paired data.
• The method enhances the utilization of complementary information from anatomical and functional MRI, leading to more precise tumor segmentation.
• The findings suggest potential for improved diagnosis and treatment evaluation of NSCLC.