Using Spatial Probability Maps to Highlight Potential Inaccuracies in Deep Learning-Based Contours: Facilitating...

Area of Science:

• Medical imaging and radiation oncology.
• Artificial intelligence in healthcare.
• Image segmentation and contouring.

Background:

• Manual contouring of organs at risk is time-consuming and variable.
• Deep learning-based delineation (DLD) offers speed and quality but is not perfect.
• Manual checking of DLD is recommended, but tools for identifying contour uncertainty are lacking.

Purpose of the Study:

• To explore the use of spatial probability maps (SPMs) for improving efficiency and reproducibility of DLD checking and correction.
• To utilize SPMs to highlight areas of greatest uncertainty in DLD contours.
• To use salivary glands as a model for this investigation.

Main Methods:

• Trained a 3D fully convolutional network on parotid and submandibular glands.
• Generated SPMs using Monte Carlo dropout (MCD).
• Enhanced MCD with Gaussian distribution (GD) over model parameters (MCD + GD) and compared with MCD, including visual inspection of SPMs.

Main Results:

• The addition of Gaussian distribution (GD) enhanced the detection of uncertainty.
• SPMs identified uncertainty in areas with lower contrast, less consistent clinical contouring, and deviations from the anatomic norm.
• The technique demonstrated effectiveness in highlighting unreliable contour areas.

Conclusions:

• Integrating uncertainty information into DLD contours is crucial for reliability assessment.
• SPMs provide a method to highlight contour uncertainty.
• SPMs can be integrated into online adaptive radiation therapy workflows.