LEARNABLE HIERARCHICAL VISUAL CONTEXTS FOR TUMOR SEGMENTATION IN COMPUTED TOMOGRAPHY IMAGES

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Automated tumor segmentation on CT scans is crucial for radiotherapy but remains challenging due to tumor variability.
• Deep learning (DL) methods struggle with diffuse boundaries and appearance variations in tumors.

Purpose of the Study:

• To introduce LeVal, a novel deep learning approach for enhanced automated tumor segmentation on CT scans.
• To improve the accuracy and discriminability of tumor segmentation in radiotherapy applications.

Main Methods:

• LeVal utilizes learnable visual query contexts (semantic and task-specific tokens) with a 3D Swin transformer encoder.
• A two-stage pretraining strategy involving self-supervised learning (SSL) on unlabeled CTs and supervised pretraining for multi-organ segmentation was employed.
• Task queries are refined via cross-attention with semantic contexts to modulate the decoder output for segmentation.

Main Results:

• LeVal consistently outperformed existing methods across four public datasets (pancreas, colon, adrenal, head-and-neck cancers).
• The method demonstrated improved embedding separation between tumor and healthy tissues, indicating enhanced discriminability.
• The approach refines attention towards tumor-relevant regions for more precise segmentation.

Conclusions:

• LeVal offers a significant advancement in automated tumor segmentation for radiotherapy.
• The proposed method enhances the accuracy and robustness of deep learning models in medical image analysis.
• Availability of code and model checkpoints will facilitate further research and application.