Assessing and testing anomaly detection for finding prostate cancer in spatially registered multi-parametric MRI

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Multi-Parametric MRI (MP-MRI) is crucial for non-invasive prostate cancer management.
• Supervised algorithms have been used to analyze MP-MRI data.
• This study explores an unsupervised approach for prostate cancer detection using MP-MRI.

Purpose of the Study:

• To apply the Reed-Xiaoli (RX) anomaly detector, an unsupervised algorithm, to prostate cancer detection using MP-MRI.
• To evaluate the performance of RX in identifying tumor voxels compared to existing methods.
• To assess the utility of RX for enhancing the analysis of spatially registered MP-MRI data.

Main Methods:

• Prospective collection of MP-MRI data (T1, T2, diffusion, DCE) from 26 patients.
• Spatial registration and formation of multi-parametric cubes from MP-MRI data.
• Application of the RX anomaly detector with various noise reduction techniques (PC filtering, regularization) and comparison with Adaptive Cosine Estimator (ACE) and quantitative color analysis using ROC curves.

Main Results:

• RX with principal component (PC) filtering (3 or 4 PCs) and regularization achieved higher Area Under the Curve (AUC) and Youden Index (YI) compared to other RX variations and reference methods.
• The best performing RX configurations yielded AUCs around 0.74 with YIs around 0.71.
• Standard errors for all measurements were below 0.020, indicating reliable results.

Conclusions:

• The Reed-Xiaoli (RX) algorithm, an unsupervised anomaly detector, was successfully applied to spatially registered MP-MRI for prostate cancer detection.
• Filtering principal components and applying regularization significantly improved RX performance.
• This unsupervised approach offers a novel method for identifying aberrant voxels indicative of prostate cancer in MP-MRI scans.