Effect of vascular lesion preprocessing on Brain Intensity AbNormality Classification Algorithm (BIANCA) white matter...

Area of Science:

• Neuroimaging
• Medical image analysis
• Cerebrovascular disease research

Background:

• White matter hyperintensity (WMH) segmentation in stroke is challenging due to overlapping lesion characteristics.
• Systematic evaluation of preprocessing impacts on WMH segmentation in multi-scanner cerebrovascular studies is lacking.

Purpose of the Study:

• To compare thresholding methods and lesion-handling strategies for WMH segmentation using BIANCA.
• To assess the impact of preprocessing decisions on segmentation accuracy in stroke populations.

Main Methods:

• Compared fixed vs. LOCATE adaptive thresholding and lesion removal, zero-intensity replacement, or inpainting.
• Utilized the BeLOVE and WMH Segmentation Challenge datasets for optimization and validation.
• Employed 5-fold cross-validation for threshold optimization and assessed segmentation accuracy and volume agreement.

Main Results:

• BIANCA with LOCATE achieved moderate WMH segmentation (Dice 0.567) and high lesion-level detection (F1 > 0.85).
• Preprocessing effects on segmentation accuracy were statistically detectable but negligible across conditions.
• Stroke lesion volume was the primary predictor of volume differences, scaling with size but remaining negligible.

Conclusions:

• BIANCA with LOCATE offers a balanced sensitivity-precision trade-off for WMH segmentation in multi-scanner cerebrovascular cohorts.
• Group-level preprocessing effects were minimal, but large lesions can distort BIANCA's classification.
• Lesion removal is a justified preprocessing step to mitigate distortions caused by large lesions.