Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Effect of Lossy Compression on 3D Medical Images Segmentation with Deep Learning

Anvar Kurmukov, Bogdan Zavolovich, Aleksandra Dalechina, Vladislav Proskurov, Boris Shirokikh

TL;DR

It is demonstrated that lossy compression up to 20 times have no negative impact on segmentation quality with deep neural networks (DNN) and the ability of DNN models trained on compressed data to predict on uncompressed data and vice versa with no quality deterioration.

Abstract

Image compression is a critical tool in decreasing the cost of storage and improving the speed of transmission over the internet. While deep learning applications for natural images widely adopts the usage of lossy compression techniques, it is not widespread for 3D medical images. Using three CT datasets (17 tasks) and one MRI dataset (3 tasks) we demonstrate that lossy compression up to 20 times have no negative impact on segmentation quality with deep neural networks (DNN). In addition, we demonstrate the ability of DNN models trained on compressed data to predict on uncompressed data and vice versa with no quality deterioration.

The Effect of Lossy Compression on 3D Medical Images Segmentation with Deep Learning

TL;DR

It is demonstrated that lossy compression up to 20 times have no negative impact on segmentation quality with deep neural networks (DNN) and the ability of DNN models trained on compressed data to predict on uncompressed data and vice versa with no quality deterioration.

Abstract

Image compression is a critical tool in decreasing the cost of storage and improving the speed of transmission over the internet. While deep learning applications for natural images widely adopts the usage of lossy compression techniques, it is not widespread for 3D medical images. Using three CT datasets (17 tasks) and one MRI dataset (3 tasks) we demonstrate that lossy compression up to 20 times have no negative impact on segmentation quality with deep neural networks (DNN). In addition, we demonstrate the ability of DNN models trained on compressed data to predict on uncompressed data and vice versa with no quality deterioration.
Paper Structure (17 sections, 3 equations, 5 figures, 2 tables)

This paper contains 17 sections, 3 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. JPEG2000
  4. Data
  5. LiTS
  6. BraTS
  7. AMOS
  8. BGPD
  9. Experiments
  10. Metrics
  11. Results
  12. Compressed images quality
  13. Effect of training on lossy compressed images
  14. Inference models trained on original data
  15. Few-shot vs lossy compressed full training
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Left: results of human assessment of image pairs with different compression ratios. Each line represents the average Spearman's rank correlation (x-axis) between two compression rates (two ends of each line, y-axis). Right: peak signal-to-noise ratio of images compressed at different rates.
  • Figure 2: Left: segmentation performance of models trained on data compressed at different rates. Right: storage memory footprint tradeoff, few-shot (uncompressed) vs full-shot (compressed).
  • Figure 3: Testing model trained on images at different compression rates on images of other comression rates. Results for AMOS dataset (left), LiTS dataset (right).
  • Figure 4: Example image used during human evaluation. Radiologists were asked to decide whether left or right image looks of higher quality (or both looks indistinguishable). CT images were clipped to abdominal, brain, or liver window respectively.
  • Figure 5: Effect of JPEG2000 compression on images from different experimental datasets.