Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-organ segmentation: a progressive exploration of learning paradigms under scarce annotation

Shiman Li, Haoran Wang, Yucong Meng, Chenxi Zhang, Zhijian Song

TL;DR

The fully supervised method is reviewed, a comprehensive and systematic elaboration of the 3 abovementioned learning paradigms in the context of multi-organ segmentation from both technical and methodological perspectives are presented, and their challenges and future trends are summarized.

Abstract

Precise delineation of multiple organs or abnormal regions in the human body from medical images plays an essential role in computer-aided diagnosis, surgical simulation, image-guided interventions, and especially in radiotherapy treatment planning. Thus, it is of great significance to explore automatic segmentation approaches, among which deep learning-based approaches have evolved rapidly and witnessed remarkable progress in multi-organ segmentation. However, obtaining an appropriately sized and fine-grained annotated dataset of multiple organs is extremely hard and expensive. Such scarce annotation limits the development of high-performance multi-organ segmentation models but promotes many annotation-efficient learning paradigms. Among these, studies on transfer learning leveraging external datasets, semi-supervised learning using unannotated datasets and partially-supervised learning integrating partially-labeled datasets have led the dominant way to break such dilemma in multi-organ segmentation. We first review the traditional fully supervised method, then present a comprehensive and systematic elaboration of the 3 abovementioned learning paradigms in the context of multi-organ segmentation from both technical and methodological perspectives, and finally summarize their challenges and future trends.

Multi-organ segmentation: a progressive exploration of learning paradigms under scarce annotation

TL;DR

The fully supervised method is reviewed, a comprehensive and systematic elaboration of the 3 abovementioned learning paradigms in the context of multi-organ segmentation from both technical and methodological perspectives are presented, and their challenges and future trends are summarized.

Abstract

Precise delineation of multiple organs or abnormal regions in the human body from medical images plays an essential role in computer-aided diagnosis, surgical simulation, image-guided interventions, and especially in radiotherapy treatment planning. Thus, it is of great significance to explore automatic segmentation approaches, among which deep learning-based approaches have evolved rapidly and witnessed remarkable progress in multi-organ segmentation. However, obtaining an appropriately sized and fine-grained annotated dataset of multiple organs is extremely hard and expensive. Such scarce annotation limits the development of high-performance multi-organ segmentation models but promotes many annotation-efficient learning paradigms. Among these, studies on transfer learning leveraging external datasets, semi-supervised learning using unannotated datasets and partially-supervised learning integrating partially-labeled datasets have led the dominant way to break such dilemma in multi-organ segmentation. We first review the traditional fully supervised method, then present a comprehensive and systematic elaboration of the 3 abovementioned learning paradigms in the context of multi-organ segmentation from both technical and methodological perspectives, and finally summarize their challenges and future trends.
Paper Structure (19 sections, 4 figures, 5 tables)

This paper contains 19 sections, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Overview of multi-organ segmentation & problem analysis
  3. Learning paradigms under annotation scarcity
  4. Leveraging external annotated datasets - Transfer learning
  5. Dimension transfer
  6. Source transfer
  7. Modality transfer
  8. Using unannotated datasets – Semi-supervised learning
  9. Pseudo label-based methods
  10. Consistency regularization-based methods
  11. Hybrid methods
  12. Fusing partially annotated datasets - Partially-supervised learning
  13. Unified training
  14. Separative training
  15. Discussion and future prospects
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Organization of this review paper. We roundly present the basic segmentation mode of multi-organ segmentation in the fully supervised learning methods and categorize the annotation-efficient learning methods into three classes to embrace the scarce annotations. Each annotation-efficient learning will be categorized and described based on its specific technical details.
  • Figure 2: The diagram of dual-stream models in the work by Valindria et al. valindria_multi-modal_2018. a) modality-specific encoder and shared decoder. b) shared encoder and modality-specific decoder. c) shared encoder and decoder. d) different streams in the encoder, shared last layers of the encoder, and modality-specific decoder.
  • Figure 3: Typical workflows semi-supervised methods. a) a basic workflow of the pseudo-labeling method, which uses the prediction of unlabeled data for pseudo label-based supervision; b) an example of consistency-based training using transformation consistency, in which the predictions from two networks are supervised by a consistent regularization; c) an example of a hybrid method using cross pseudo supervised, which construct a regularization by exchanging pseudo labels.
  • Figure 4: Frameworks of partially-supervised learning in multi-organ segmentation. The first row presents two kinds of unified training frameworks, where the same multi-organ segmentation network is trained jointly with different datasets. The second row presents three kinds of separative training, where different data flows through slightly different multi-organ segmentation network which will change with different classes.