Boundary guidance network for medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Oncology

Background:

• Accurate bladder tumor segmentation is vital for patient treatment and prognosis.
• Current deep learning methods primarily focus on CT/MRI, neglecting cystoscopy findings.
• Diverse tumor morphology and fuzzy boundaries present significant segmentation challenges.

Purpose of the Study:

• To propose a novel boundary-guided network for accurate bladder tumor segmentation from cystoscopy images.
• To enhance tumor feature extraction by combining CNN local features and Parallel ViT long-range dependencies.
• To improve segmentation accuracy and boundary preservation in medical imaging.

Main Methods:

• Developed a boundary guidance network integrating CNNs and Parallel ViT.
• Introduced a boundary extraction module to guide the decoding process.
• Implemented foreground-background dual-channel decoding via a boundary integration module.

Main Results:

• Achieved high performance on a new cystoscopic bladder tumor dataset (BTD) with IoU of 91.3%, Hausdorff Distance of 10.43, mAP of 85.3%, and F1 score of 94.8%.
• Demonstrated superior segmentation accuracy and boundary information retention compared to state-of-the-art methods.
• Validated effectiveness across BTD and three other public medical image datasets.

Conclusions:

• The proposed boundary-guided network effectively segments bladder tumors from cystoscopy images.
• The integration of local and long-range features, along with boundary guidance, enhances segmentation performance.
• This model shows significant promise for improving bladder cancer diagnosis and treatment planning.