Multi-sequence brain tumor segmentation boosted by deep semantic features.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning (DL) based brain tumor segmentation requires high intra-class consistency in learned image features.
• Heterogeneous brain tumors cause diverse image gray values, leading to low intra-class consistency and inefficient segmentation.
• This diversity complicates the projection of image features to their corresponding semantic labels.

Purpose of the Study:

• To address the challenge of low intra-class consistency in brain tumor image features.
• To enhance the projection of image features to semantic labels for accurate brain tumor segmentation.
• To improve the overall performance of DL-based brain tumor segmentation.

Main Methods:

• A novel DL-based method incorporating a semantic feature module (SFM) for brain tumor segmentation is proposed.
• The SFM consolidates image features with semantic information and enhances intra-class consistency.
• Deep semantic vectors are derived as prototypes to re-encode image features, reducing diversity and enriching semantic information.

Main Results:

• The proposed method was evaluated on the BraTS2022 dataset, comprising multi-sequence MR images from 1251 patients.
• The SFM significantly improved brain tumor sub-region segmentation accuracy compared to state-of-the-art methods (statistically significant).
• Ablation studies revealed up to an 11% improvement in segmentation accuracy (Dice index) with the SFM.

Conclusions:

• Low intra-class consistency of learned image features negatively impacts DL-based segmentation performance.
• The proposed SFM effectively enhances intra-class consistency using high-level semantic information.
• This enhancement leads to more accurate projection of image features to semantic labels and improved segmentation.