Unsupervised Deep Learning for Blood-Brain Barrier Leakage Detection in Diffuse Glioma Using Dynamic...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Blood-brain barrier (BBB) leakage is a critical indicator in various neurological conditions.
• Conventional methods for detecting BBB leakage using dynamic contrast-enhanced MRI (DCE-MRI) often require complex pharmacokinetic models and arterial input function (AIF) estimation.
• These traditional approaches can be time-consuming and may limit generalizability.

Purpose of the Study:

• To develop and validate an unsupervised deep learning framework for robust and generalizable blood-brain barrier leakage detection.
• To eliminate the need for pharmacokinetic modeling and AIF estimation in DCE-MRI analysis.
• To provide a more efficient and accessible method for identifying BBB abnormalities.

Main Methods:

• A retrospective study utilizing DCE-MRI data from 274 patients (2010-2020).
• An autoencoder-based anomaly detection approach was employed to identify voxel-wise time-series abnormal signals through reconstruction residuals.
• Residual leakage signals (RLSs) were isolated and compared against traditional Ktrans metrics, with generalizability assessed on subsampled data and IDH mutation status prediction.

Main Results:

• The developed RLS maps demonstrated high structural similarity (0.91 ± 0.02) and correlation (r = 0.56) with Ktrans.
• RLS maps exhibited superior performance on subsampled data compared to Ktrans maps, showing better correlation (0.89 vs 0.72), higher SNR (33.09 dB vs 28.94 dB), and improved structural similarity (0.92 vs 0.87).
• RLS maps significantly outperformed Ktrans maps in predicting IDH mutation status (AUC, 0.87 vs 0.81; P = .02).

Conclusions:

• The unsupervised deep learning framework effectively detects BBB leakage using DCE-MRI.
• This novel approach obviates the necessity for pharmacokinetic models and AIF estimation, simplifying the analysis.
• The method shows strong generalizability and superior performance in predicting clinical outcomes like IDH mutation status.