Contrast quality control for segmentation task based on deep learning models-Application to stroke lesion in CT...

Area of Science:

• Medical Imaging
• Machine Learning
• Radiology

Background:

• Medical imaging is vital for stroke management, with machine learning (ML) increasingly used for lesion segmentation.
• A key challenge in subacute stroke lesion segmentation using computed tomography (CT) is insufficient image contrast.

Purpose of the Study:

• To develop and validate a method for assessing image contrast quality in CT datasets for stroke lesion segmentation.
• To identify a critical contrast threshold below which ML models fail to learn effectively.

Main Methods:

• A machine learning model was trained to assess contrast quality using Fisher's ratio to measure lesion-background contrast.
• Performance, graphical, and clustering analyses were employed to determine the critical contrast threshold.
• The method was applied to brain lesion segmentation in CT imaging.

Main Results:

• A Fisher's ratio threshold of 0.05 was identified for adequate contrast.
• Training a new model without low-contrast images (below the threshold) reduced training data by 40% and initial training time by nearly 30%.
• The model trained on improved data performed comparably to a model trained on the full dataset when tested on external data.

Conclusions:

• The proposed method effectively identifies and excludes low-contrast images, improving dataset design and accelerating ML model training for stroke lesion segmentation.
• This approach has the potential to be adapted for other medical imaging segmentation tasks beyond stroke.