RACLAHE

RACLAHE enhances contrast in T2-Weighted Magnetic Resonance Imaging (MRI) scans using Region-Adaptive Contrast Limited Adaptive Histogram Equalization to improve convolutional neural network (CNN) segmentation and delineation of the prostate and prostatic zones.

Key Features:

• Region-Adaptive CLAHE: Applies Region-Adaptive Contrast Limited Adaptive Histogram Equalization to adaptively enhance contrast across different regions of T2-Weighted MRI scans.
• Targeted anatomy: Specifically designed to improve segmentation and delineation of the prostate and prostatic zones on T2-Weighted MRI.
• CNN performance improvement: Demonstrates consistent improvement in CNN-based segmentation accuracy across tested models.
• Cross-model evaluation: Evaluated against four other image enhancement techniques to benchmark performance.
• Multi-architecture testing: Tested on five CNN architectures: U-Net, U-Net++, U-Net3+, ResU-net, and USE-NET.
• Quantified gains: Reported mean Dice Score increases ranging from 3% to 9% for various prostatic regions.
• Low inter-model variability: Exhibits minimal variability in performance across different CNN architectures.
• Feature-driven interpretability: Implements a feature-driven approach that quantitatively and qualitatively relates saliency maps to ground truth probability maps to explain how enhancements influence predictions.

Scientific Applications:

• Prostate segmentation: Improves automated segmentation of the prostate on T2-Weighted MRI for research and clinical workflows.
• Prostatic zone delineation: Enhances delineation of prostatic zones to support region-specific analysis.
• Model benchmarking: Serves as an image-preprocessing benchmark when comparing CNN segmentation architectures (U-Net, U-Net++, U-Net3+, ResU-net, USE-NET).
• Interpretability studies: Facilitates quantitative and qualitative studies linking saliency maps to ground truth probability maps to increase transparency of CNN predictions.
• Clinical decision support: Supports tasks relevant to prostate cancer diagnosis and treatment planning by improving segmentation accuracy on T2-Weighted MRI.

Methodology:

Applies Region-Adaptive Contrast Limited Adaptive Histogram Equalization to T2-Weighted MRI; evaluates segmentation performance against four other enhancement methods using U-Net, U-Net++, U-Net3+, ResU-net, and USE-NET architectures; quantifies results with Dice Score and performs quantitative and qualitative feature-driven analysis comparing saliency maps to ground truth probability maps.