Source-linked AI summary
Self-Supervised Learning for Cardiac MR Image Segmentation by Anatomical Position Prediction
Wenjia Bai, Chen Chen, Giacomo Tarroni, Jinming Duan, Florian Guitton, Steffen E. Petersen, Yike Guo, Paul M. Matthews, Daniel Rueckert
TL;DR
Cardiac MR segmentation requires costly expert annotations, motivating feature learning from unannotated scans. The paper predicts anatomical positions as a self-supervised task and transfers the learned features to segmentation, outperforming or matching a scratch-trained U-net, especially with little data.
Problem
Cardiac MR segmentation methods rely on large amounts of manually annotated data, but such annotations are expensive and limited by expert resources.
Method
The method learns features by predicting anatomical positions automatically defined by cardiac chamber view planes, then transfers them to cardiac MR segmentation.
Results
Self-supervised learning outperforms a scratch-trained U-net at small data settings and achieves comparable performance at large data settings on short-axis and long-axis segmentation.
Takeaways & Limitations
Anatomical positions from standard cardiac MR scans provide a strong supervisory signal for improving segmentation accuracy without extra manual annotation.
Abstract
from arXiv · showhide
In the recent years, convolutional neural networks have transformed the field of medical image analysis due to their capacity to learn discriminative image features for a variety of classification and regression tasks. However, successfully learning these features requires a large amount of manually annotated data, which is expensive to acquire and limited by the available resources of expert image analysts. Therefore, unsupervised, weakly-supervised and self-supervised feature learning techniques receive a lot of attention, which aim to utilise the vast amount of available data, while at the same time avoid or substantially reduce the effort of manual annotation. In this paper, we propose a novel way for training a cardiac MR image segmentation network, in which features are learnt in a self-supervised manner by predicting anatomical positions. The anatomical positions serve as a supervisory signal and do not require extra manual annotation. We demonstrate that this seemingly simple task provides a strong signal for feature learning and with self-supervised learning, we achieve a high segmentation accuracy that is better than or comparable to a U-net trained from scratch, especially at a small data setting. When only five annotated subjects are available, the proposed method improves the mean Dice metric from 0.811 to 0.852 for short-axis image segmentation, compared to the baseline U-net.
1 Introduction
Cardiac MR segmentation benefits from CNNs but depends on costly, limited expert annotations. The paper focuses on self-supervised feature learning from unannotated data to reduce annotation requirements.
- Cardiac MR segmentation produces quantitative phenotypes used as biomarkers for diagnosing and managing cardiovascular diseases.
- Most successful cardiac MR segmentation methods are fully supervised and require large amounts of annotated data to learn image features.Annotations are expensive to acquire and limited by expert image-analysis resources.
- The paper uses anatomical positions automatically defined by cardiac chamber view planes as a novel self-supervised pretext task.The chamber view-plane information is available from standard cardiac MR scans without extra manual annotation.
- Self-supervised learning is evaluated on short-axis and long-axis segmentation and provides a strong accuracy boost, especially with few training subjects.
- Self-supervised learning formulates a pretext task from unannotated data for feature learning.
2 Methods
The method predicts anatomical positions defined by cardiac MR view planes, then transfers the learned features to anatomical-structure segmentation. It compares decoder-only, all-weight, and multi-task transfer strategies.
- Cardiac MR view planes:: Cardiac MR scans include short-axis, long-axis 2Ch, 4Ch, and 3Ch views acquired at different angulated planes.These views evaluate different heart regions, including ventricular cross-sections, ventricular walls, and atrial chambers.
- Cardiac MR view planes:: The method uses relative short-axis and long-axis orientation and view-plane-defined anatomical positions as a self-supervised pretext task.This addresses prior approaches that consider the views separately and disregard their relative orientation.
- Cardiac MR view planes:: Nine anatomical positions are defined along intersecting chamber-view lines, including the intersection and boxes arranged left, right, anterior, and posterior.
- Self-Supervised Learning (SSL):: The pretext network is a 10-way segmentation model that predicts nine anatomical-position boxes plus background.A standard U-net with encoder, decoder, skip connections, and task head is trained with cross-entropy.
- Self-Supervised Learning (SSL):: Learned pretext features transfer to segmentation of structures such as the LV cavity, myocardium, and RV cavity by replacing the task head.The new head is a 1×1 convolution with K-channel output, where K is the number of classes.
- Transfer Learning:: Transfer learning freezes the encoder with decoder-only tuning, finetunes all weights, or uses multi-task finetuning with two task heads.
- Transfer Learning:: Multi-task finetuning combines pretext and segmentation losses as L(θ) = Ltask1(x, y1|θ) + β · Ltask2(x, y2|θ).The labels represent anatomical positions for task 1 and expert-annotated anatomical structures for task 2; β weights the second loss.
3 Experiments and Results
Experiments evaluate self-supervised transfer for short-axis and long-axis cardiac MR segmentation. SSL methods generally outperform U-net-scratch with limited training data, while differences narrow with more data.
- Experimental setup: 3,825 subjects supplied short-axis and long-axis images for self-supervised learning, while transfer learning used 200 subjects with manual annotations.Bounding boxes were automatically placed at nine anatomical positions on short-axis and long-axis 4Ch images.
- Implementation: Training used Adam with a 0.001 learning rate, online rotation and scaling augmentation, and alternating multi-task optimization.The baseline U-net-scratch used the same setting but random initialization.
- Short-axis image segmentation: SSL+Decoder achieved accuracy comparable to U-net-scratch despite tuning only the decoder, indicating transferable encoder features.SSL+All and SSL+MultiTask tune all weights during transfer learning.
- Short-axis image segmentation: SSL+MultiTask generally achieved the best short-axis segmentation, especially when the number of training subjects was small.The advantage was observed when all network weights were tuned.
- Short-axis image segmentation: For short-axis segmentation, SSL+MultiTask outperformed U-net-scratch at n ≤10, whereas their performances became close at n ≥50.At n = 1, U-net-scratch failed completely, while SSL+MultiTask segmented parts of the LV and myocardium; at n = 5 or 10, it avoided RV under-segmentation errors.
- Long-axis image segmentation: SSL+MultiTask generally produced better long-axis segmentations than U-net-scratch with limited training data.Table 2 reports mean Dice overlap metrics, and Figure 5 provides exemplar segmentations across increasing training-set sizes.
4 Conclusions
The paper proposes anatomical position prediction as a self-supervised pretext task for cardiac MR segmentation. It outperforms U-net-scratch in small-data settings and performs comparably with larger datasets.
- 4 Conclusions: The method uses anatomical position prediction to learn features for cardiac MR image segmentation without requiring extra manual annotation.The pretext task provides the supervisory signal for self-supervised learning.
- 4 Conclusions: Self-supervised learning outperforms a standard U-net trained from scratch at small data settings and achieves comparable performance at large data settings.The conclusion is supported by experiments on both short-axis and long-axis segmentation tasks.
- 4 Conclusions: Future work will explore other anatomically meaningful pretext tasks to increase data efficiency in medical imaging applications.