Incorporating minimal user input into deep learning based image segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer vision

Background:

• Computer-assisted image segmentation aids clinicians in faster border delineation with reduced variability.
• Convolutional neural networks (CNNs) are increasingly utilized for automated image segmentation tasks.

Purpose of the Study:

• To develop and evaluate a technique for supervising CNNs using observer inputs to enhance segmentation accuracy.
• To investigate the impact of sparse surface points as additional input for CNN-based image segmentation.
• To compare the performance of U-Net and a novel dual-input path (DIP) U-Net architecture for supervised prostate MRI segmentation.

Main Methods:

• Implemented a novel supervision technique involving sparse surface points as input to CNNs.
• Tested the technique using U-Net and a new DIP U-Net architecture for prostate segmentation on MRI.
• Compared the supervised segmentation results against fully automatic segmentation and inter-expert manual segmentation variability.

Main Results:

• The observer-input supervision technique significantly increased segmentation accuracy for both U-Net and DIP U-Net compared to fully automatic segmentation.
• The DIP U-Net architecture demonstrated superior performance over the standard U-Net in supervised image segmentation.
• Segmentation performance comparable to manual segmentation was achieved using approximately 15 to 20 selected surface points.

Conclusions:

• Supervising CNNs with minimal clinician interaction, such as sparse surface points, is an effective strategy to improve segmentation accuracy.
• The DIP U-Net architecture shows promise for enhanced supervised medical image segmentation.
• This approach offers a viable method to achieve segmentation accuracy comparable to expert manual delineation, reducing inter-observer variability.