Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UNGT: Ultrasound Nasogastric Tube Dataset for Medical Image Analysis

Zhaoshan Liu, Chau Hung Lee, Qiujie Lv, Nicole Kessa Wee, Lei Shen

Abstract

We develop a novel ultrasound nasogastric tube (UNGT) dataset to address the lack of public nasogastric tube datasets. The UNGT dataset includes 493 images gathered from 110 patients with an average image resolution of approximately 879 $\times$ 583. Four structures, encompassing the liver, stomach, tube, and pancreas, are precisely annotated. Besides, we propose a semi-supervised adaptive-weighting aggregation medical segmenter to address data limitation and imbalance concurrently. The introduced adaptive weighting approach tackles the severe unbalanced challenge by regulating the loss across varying categories as training proceeds. The presented multiscale attention aggregation block bolsters the feature representation by integrating local and global contextual information. With these, the proposed AAMS can emphasize sparse or small structures and feature enhanced representation ability. We perform extensive segmentation experiments on our UNGT dataset, and the results show that AAMS outperforms existing state-of-the-art approaches to varying extents. In addition, we conduct comprehensive classification experiments across varying state-of-the-art methods and compare their performance. The dataset and code will be available upon publication at https://github.com/NUS-Tim/UNGT.

UNGT: Ultrasound Nasogastric Tube Dataset for Medical Image Analysis

Abstract

We develop a novel ultrasound nasogastric tube (UNGT) dataset to address the lack of public nasogastric tube datasets. The UNGT dataset includes 493 images gathered from 110 patients with an average image resolution of approximately 879 583. Four structures, encompassing the liver, stomach, tube, and pancreas, are precisely annotated. Besides, we propose a semi-supervised adaptive-weighting aggregation medical segmenter to address data limitation and imbalance concurrently. The introduced adaptive weighting approach tackles the severe unbalanced challenge by regulating the loss across varying categories as training proceeds. The presented multiscale attention aggregation block bolsters the feature representation by integrating local and global contextual information. With these, the proposed AAMS can emphasize sparse or small structures and feature enhanced representation ability. We perform extensive segmentation experiments on our UNGT dataset, and the results show that AAMS outperforms existing state-of-the-art approaches to varying extents. In addition, we conduct comprehensive classification experiments across varying state-of-the-art methods and compare their performance. The dataset and code will be available upon publication at https://github.com/NUS-Tim/UNGT.

Paper Structure

This paper contains 17 sections, 5 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Medical Ultrasound Dataset
  4. Medical Image Analysis
  5. Methods
  6. Data Acquisition and Characteristics
  7. Model Architecture Design
  8. Experiments
  9. Dataset Preprocessing
  10. Configuration and Evaluation
  11. Results and Analysis
  12. Segmentation
  13. Classification
  14. Ablation and Generalization
  15. Ablation Study
  16. ...and 2 more sections

Figures (7)

  • Figure 1: Example images and characteristics analysis of the developed UNGT dataset. Red, green, orange, and blue regions represent the liver, stomach, tube, and pancreas areas. Com. A and Com. B represents the two components. (a) Example images in UNGT, (b) image counts per patient, (c) grey level histogram without zero level, (d) image counts per structure, and (e) pixel counts per structure.
  • Figure 2: Architecture of the developed AAMS. It comprises an ADW scheme, an MAA block, the OAA, and the RRE blocks. The loss function of AAMS includes fusion loss $L_{f}$, sensitivity loss $L_{s}$, and unsupervised loss $L_{u}$. The ADW scheme calculates each class's normalized loss weight $\omega'$. The MAA block is mainly composed of a local excitation (LEX) block, a global excitation (GEX) block, and a multi-layer perceptron (MLP) block. $\oplus$ and $\times$ denote the exclusive OR and multiplication operations. L, S, T, and P stand for the liver, stomach, tube, and pancreas. HAAP and VAAP represent horizontal and vertical adaptive average pooling layers. DwConv, SpConv, and PwConv present depthwise, stripe, and pointwise convolution layers. BN stands for batch normalization layer. (a) AAMS architecture, (b) ADW scheme, and (c) MAA block.
  • Figure 3: Predicted and GT masks across AAMS and SOTA approaches on the introduced UNGT dataset. GT stands for ground truth.
  • Figure 4: Cumulative distribution function of varying structures regarding DSC across different methods on the developed UNGT dataset. The horizontal red dotted lines depict the proportion of the GT mask without the corresponding structures in percentage. (a) Liver, (b) stomach, (c) tube, and (d) pancreas.
  • Figure 5: Bland-Altman analysis of AAMS across varying structures on the developed UNGT dataset. LOA demonstrates the limits of agreement. (a) Liver, (b) stomach, (c) tube, and (d) pancreas.
  • ...and 2 more figures