Semantic segmentation to identify bladder layers from H&E Images.

Area of Science:

• Digital pathology
• Computational oncology
• Medical image analysis

Background:

• Accurate identification of bladder layers is crucial for diagnosing and prognosticating bladder cancer.
• Current methods rely on manual annotation, which can be time-consuming.
• Hematoxylin and eosin (H&E) stained slides are standard for bladder biopsy analysis.

Purpose of the Study:

• To develop and evaluate a multi-class image segmentation method for bladder biopsies.
• To accurately identify key bladder layers (urothelium, lamina propria, muscularis propria) and pathological regions (red blood cells, cauterized/inflamed tissue).
• To assess the potential of AI in assisting pathologists with bladder cancer diagnosis.

Main Methods:

• A U-Net convolutional neural network architecture was employed for image segmentation.
• Model performance was optimized by testing variations in network depth (8, 10, 12 layers) and weight initializers (He uniform, He normal, Glorot uniform, Glorot normal).
• A seven-fold cross-validation approach was used on a dataset of 39 whole slide images from T1 bladder biopsies.

Main Results:

• The optimal model configuration was a 12-layer U-Net with He normal initialization.
• Pathologist review of AI-segmented slides showed high accuracy (average scores of 8.93/10 and 8.87/10 on independent datasets).
• The AI model significantly reduced slide review time, with an average of 1.53 minutes per slide.

Conclusions:

• The developed AI model demonstrates high accuracy in segmenting bladder tissue layers.
• The method has the potential to expedite the slide annotation process for pathologists.
• Further development could enhance diagnostic capabilities for T1 bladder cancer.