Reproduce “Mono2D: A Trainable Monogenic Layer for Robust Knee Cartilage Segmentation on Out-of-Distribution 2D Ultrasound Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alvin Kimbowa, Arjun Parmar, Maziar Badii, David Liu, Matthew Harkey, Ilker Hacihaliloglu
A trainable monogenic layer, Mono2D, extracts multi-scale, contrast- and intensity-invariant local phase features to enhance deep learning models' robustness for medical image segmentation on out-of-distribution data. This layer achieved an average Dice score of 94.21% on knee cartilage ultrasound and 75.05% on multi-site prostate MRI, outperforming other single-source domain generalization methods.
Key insights
- Trainable multi-scale local phase features inherently offer superior robustness to intensity and contrast variations compared to traditional methods, effectively mitigating domain shift in medical images.
- Constraining Log-Gabor filter parameters with sigmoid functions during optimization allows for stable and effective learning of optimal bandpass filters for feature extraction.
- Combining both local phase and phase asymmetry from the monogenic signal provides a more comprehensive and resilient representation of anatomical structures, enhancing segmentation accuracy across domains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.