Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Efficient and Accurate CT Segmentation via Edge-Preserving Probabilistic Downsampling

Shahzad Ali, Yu Rim Lee, Soo Young Park, Won Young Tak, Soon Ki Jung

TL;DR

Edge-preserving Probabilistic Downsampling (EPD) utilizes class uncertainty within a local window to produce soft labels, with the window size dictating the downsampling factor, which surpasses bilinear interpolation in image downsampling, enhancing overall performance.

Abstract

Downsampling images and labels, often necessitated by limited resources or to expedite network training, leads to the loss of small objects and thin boundaries. This undermines the segmentation network's capacity to interpret images accurately and predict detailed labels, resulting in diminished performance compared to processing at original resolutions. This situation exemplifies the trade-off between efficiency and accuracy, with higher downsampling factors further impairing segmentation outcomes. Preserving information during downsampling is especially critical for medical image segmentation tasks. To tackle this challenge, we introduce a novel method named Edge-preserving Probabilistic Downsampling (EPD). It utilizes class uncertainty within a local window to produce soft labels, with the window size dictating the downsampling factor. This enables a network to produce quality predictions at low resolutions. Beyond preserving edge details more effectively than conventional nearest-neighbor downsampling, employing a similar algorithm for images, it surpasses bilinear interpolation in image downsampling, enhancing overall performance. Our method significantly improved Intersection over Union (IoU) to 2.85%, 8.65%, and 11.89% when downsampling data to 1/2, 1/4, and 1/8, respectively, compared to conventional interpolation methods.

Towards Efficient and Accurate CT Segmentation via Edge-Preserving Probabilistic Downsampling

TL;DR

Edge-preserving Probabilistic Downsampling (EPD) utilizes class uncertainty within a local window to produce soft labels, with the window size dictating the downsampling factor, which surpasses bilinear interpolation in image downsampling, enhancing overall performance.

Abstract

Downsampling images and labels, often necessitated by limited resources or to expedite network training, leads to the loss of small objects and thin boundaries. This undermines the segmentation network's capacity to interpret images accurately and predict detailed labels, resulting in diminished performance compared to processing at original resolutions. This situation exemplifies the trade-off between efficiency and accuracy, with higher downsampling factors further impairing segmentation outcomes. Preserving information during downsampling is especially critical for medical image segmentation tasks. To tackle this challenge, we introduce a novel method named Edge-preserving Probabilistic Downsampling (EPD). It utilizes class uncertainty within a local window to produce soft labels, with the window size dictating the downsampling factor. This enables a network to produce quality predictions at low resolutions. Beyond preserving edge details more effectively than conventional nearest-neighbor downsampling, employing a similar algorithm for images, it surpasses bilinear interpolation in image downsampling, enhancing overall performance. Our method significantly improved Intersection over Union (IoU) to 2.85%, 8.65%, and 11.89% when downsampling data to 1/2, 1/4, and 1/8, respectively, compared to conventional interpolation methods.
Paper Structure (14 sections, 1 equation, 4 figures, 1 table, 1 algorithm)

This paper contains 14 sections, 1 equation, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. Problem Definition
  4. Downsampling Algorithm
  5. For Labels
  6. For Images
  7. Training a Multi-class Segmentation Network
  8. Experiments
  9. Dataset
  10. Preprocessing
  11. Evaluation Metrics
  12. Implementation Details
  13. Results and Discussion
  14. Conclusions

Figures (4)

  • Figure 1: EPD exhibits an inverted pyramid-like structure with the highest and the lowest resolutions at the base and apex, denoted as $S_0$ and $S_{d}$, respectively. Here, $n=9$, each level halves the preceding resolution beginning from the original resolution of ${512\times512}$.
  • Figure 2: Proposed downsampling for an original-resolution label. (a) For simplicity, input label $S_0$ is binary, and the probability of foreground class $p_1$ is calculated. (b) $d=1$ downsamples the input resolution to half. (c) $d=2$ downsamples the input resolution to quarter. The downsampling effect is prominent for windows containing edge pixels.
  • Figure 3: Application of EPD to the high-resolution inputs of $512\times512$. The conventional method employs bilinear and nearest-neighbor interpolation to downsample CT slices and labels, respectively. The window in yellow on the slices and labels is zoomed-in in the subsequent column.
  • Figure 4: Probability discrepancy between the network prediction and target labels. These randomly selected CT slices belong to six marked locations in patients with varying genders and ages. The highest and lowest discrepancy is observed for the MSC and IMAT classes. Best viewed in color.