Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mono2D: A Trainable Monogenic Layer for Robust Knee Cartilage Segmentation on Out-of-Distribution 2D Ultrasound Data

Alvin Kimbowa, Arjun Parmar, Maziar Badii, David Liu, Matthew Harkey, Ilker Hacihaliloglu

TL;DR

This work tackles domain shifts in knee cartilage segmentation from 2D ultrasound by introducing Mono2D, a trainable monogenic layer that learns multi-scale log-Gabor bandwidths to extract local phase features invariant to intensity and contrast. Integrated before the first layer of a lightweight segmentation network, Mono2D is trained jointly with the backbone and evaluated on multi-domain knee ultrasound datasets and multi-site prostate MRI data, where it outperforms six SSDG baselines in Dice and MASD. The results indicate strong cross-domain generalization and transferability across modalities, suggesting practical benefits for on-device, robust medical image segmentation. The study also highlights avenues for further enhancement of phase-based features and deeper integration in segmentation architectures.

Abstract

Automated knee cartilage segmentation using point-of-care ultrasound devices and deep-learning networks has the potential to enhance the management of knee osteoarthritis. However, segmentation algorithms often struggle with domain shifts caused by variations in ultrasound devices and acquisition parameters, limiting their generalizability. In this paper, we propose Mono2D, a monogenic layer that extracts multi-scale, contrast- and intensity-invariant local phase features using trainable bandpass quadrature filters. This layer mitigates domain shifts, improving generalization to out-of-distribution domains. Mono2D is integrated before the first layer of a segmentation network, and its parameters jointly trained alongside the network's parameters. We evaluated Mono2D on a multi-domain 2D ultrasound knee cartilage dataset for single-source domain generalization (SSDG). Our results demonstrate that Mono2D outperforms other SSDG methods in terms of Dice score and mean average surface distance. To further assess its generalizability, we evaluate Mono2D on a multi-site prostate MRI dataset, where it continues to outperform other SSDG methods, highlighting its potential to improve domain generalization in medical imaging. Nevertheless, further evaluation on diverse datasets is still necessary to assess its clinical utility.

Mono2D: A Trainable Monogenic Layer for Robust Knee Cartilage Segmentation on Out-of-Distribution 2D Ultrasound Data

TL;DR

This work tackles domain shifts in knee cartilage segmentation from 2D ultrasound by introducing Mono2D, a trainable monogenic layer that learns multi-scale log-Gabor bandwidths to extract local phase features invariant to intensity and contrast. Integrated before the first layer of a lightweight segmentation network, Mono2D is trained jointly with the backbone and evaluated on multi-domain knee ultrasound datasets and multi-site prostate MRI data, where it outperforms six SSDG baselines in Dice and MASD. The results indicate strong cross-domain generalization and transferability across modalities, suggesting practical benefits for on-device, robust medical image segmentation. The study also highlights avenues for further enhancement of phase-based features and deeper integration in segmentation architectures.

Abstract

Automated knee cartilage segmentation using point-of-care ultrasound devices and deep-learning networks has the potential to enhance the management of knee osteoarthritis. However, segmentation algorithms often struggle with domain shifts caused by variations in ultrasound devices and acquisition parameters, limiting their generalizability. In this paper, we propose Mono2D, a monogenic layer that extracts multi-scale, contrast- and intensity-invariant local phase features using trainable bandpass quadrature filters. This layer mitigates domain shifts, improving generalization to out-of-distribution domains. Mono2D is integrated before the first layer of a segmentation network, and its parameters jointly trained alongside the network's parameters. We evaluated Mono2D on a multi-domain 2D ultrasound knee cartilage dataset for single-source domain generalization (SSDG). Our results demonstrate that Mono2D outperforms other SSDG methods in terms of Dice score and mean average surface distance. To further assess its generalizability, we evaluate Mono2D on a multi-site prostate MRI dataset, where it continues to outperform other SSDG methods, highlighting its potential to improve domain generalization in medical imaging. Nevertheless, further evaluation on diverse datasets is still necessary to assess its clinical utility.

Paper Structure

This paper contains 12 sections, 6 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Mono2D layer
  4. Background
  5. Layer definition
  6. Layer training
  7. Experimental Setup
  8. Datasets
  9. Experiments
  10. Implementation details
  11. Results and Discussion
  12. Conclusion

Figures (4)

  • Figure 1: Left: The domain shift in the original datasets (top histogram) can be mitigated using local phase information (bottom histogram). Right: The relatively unclear cartilage structure in the original images (top row) is uniformly enhanced in the local phase images across all datasets (bottom row).
  • Figure 2: Approach overview: The Mono2D layer is placed before a neural network. The layer consists of a fast Fourier Transform (FFT), followed by a low-pass filter (LPF), log-Gabor filter (LGF), Riesz kernels $R_x$ and $R_y$, and inverse FFT (IFFT).
  • Figure 3: Sample qualitative results on unseen out-of-distribution images. The green contour indicates the expert label, and the red contour indicates the model prediction.
  • Figure 4: Sample qualitative results on unseen out-of-distribution images. The green contour indicates the expert label, and the red contour indicates the model prediction.