Content and shape attention network for bladder wall and cancer segmentation in MRIs

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate segmentation of bladder wall and cancer is crucial for predicting muscle-invasive status preoperatively.
• Challenges include complex backgrounds, diverse bladder shapes, and indistinct boundaries, hindering segmentation accuracy.

Purpose of the Study:

• To develop a deep network for enhanced bladder wall and cancer segmentation.
• To address segmentation challenges like complex backgrounds and weak boundaries.

Main Methods:

• Proposed a deep network integrating a content attention module (Attention U-Net) and a shape attention module (Spatial Transform Network).
• Attention U-Net emphasizes salient image features for segmentation.
• Shape attention module incorporates shape priors for closed bladder wall segmentation.

Main Results:

• The proposed model demonstrated competitive performance against existing methods.
• Achieved mean Dice Similarity Coefficients (DSCs) of 0.80 for bladder wall and 0.84 for cancer via 5-fold cross-validation.
• Effectively mitigated issues related to complex backgrounds and weak boundaries in segmentation.

Conclusions:

• The developed deep network offers an effective solution for bladder cancer and wall segmentation.
• The attention-based approach improves segmentation accuracy and reliability for clinical prediction.
• This method shows promise in overcoming common segmentation difficulties in medical imaging.