A sensitivity analysis of probability maps in deep-learning-based anatomical segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning segmentation models often implicitly learn structure presence from training data prominence.
• This implicit learning, common in natural language processing, is frequently overlooked in medical image segmentation.
• Class imbalance in datasets can lead to suboptimal segmentation performance if not properly addressed.

Purpose of the Study:

• To demonstrate the critical role of class imbalance in deep learning-based segmentation.
• To recommend adjustments to the neural network optimization objective for improved segmentation.
• To analyze the impact of structure prominence on segmentation model thresholds.

Main Methods:

• Trained 2D U-Net models for segmenting 10 head and neck structures from the Head-Neck-Radiomics-HN1 dataset.
• Determined optimal segmentation thresholds using Dice scores on validation data.
• Defined and analyzed a structure prominence measure and its effect on optimal thresholds.
• Evaluated the use of a 2D Dice objective alongside binary cross-entropy.

Main Results:

• Conventional 0.5-thresholding led to significant decreases in perceived model performance.
• Threshold perturbations of ±0.05 caused a median Dice score reduction of 11.8%.
• A weak correlation was observed between training dataset prominence and optimal thresholds.
• Optimizing with the 2D Dice score reduced thresholding variability but did not consistently yield superior segmentation.

Conclusions:

• Postprocessing procedures in deep learning-based contouring are crucial for performance enhancement.
• A mixed objective function combining binary cross-entropy and the 2D Dice score is recommended for intensity-based postprocessing.