Stochastic tissue window normalization of deep learning on computed tomography.

Area of Science:

• Medical imaging analysis
• Deep learning in radiology
• Computed tomography (CT) applications

Background:

• Tissue window filtering, particularly soft tissue windows, is common in deep learning for CT image analysis to enhance training.
• The generalizability of models trained with fixed tissue windows is questionable due to variations in CT acquisition parameters and patient physiology.

Purpose of the Study:

• To evaluate the effectiveness of soft tissue window normalization versus no normalization for CT image analysis.
• To propose and assess a novel stochastic tissue window normalization (SWN) method to improve model generalizability.

Main Methods:

• Comparison of standard soft tissue windowing, no windowing, and the proposed SWN method on multiorgan segmentation tasks.
• Utilized 80 training and 453 validation/testing scans from six diverse datasets covering same/different scanner and pathology scenarios.
• Employed a standard 2D U-Net architecture for segmentation evaluation.

Main Results:

• Traditional soft tissue windowing and non-windowed approaches performed better only when training and testing data originated from the same scanner and population.
• The proposed SWN method demonstrated superior generalizability across scenarios with different CT contrasts and pathologies.
• Statistical analyses confirmed the improved performance of SWN on diverse datasets.

Conclusions:

• Fixed tissue window normalization may limit the generalizability of deep learning models in CT image analysis.
• Stochastic tissue window normalization (SWN) offers a robust strategy to enhance model generalizability across varied CT acquisition and patient conditions.
• SWN maintains specificity for abdominal organs while improving performance on unseen data.