Source-linked AI summary
Semi-Supervised Brain Lesion Segmentation with an Adapted Mean Teacher Model
Wenhui Cui, Yanlin Liu, Yuxing Li, Menghao Guo, Yiming Li, Xiuli Li, Tianle Wang, Xiangzhu Zeng, Chuyang Ye
TL;DR
Limited, costly annotations constrain CNN-based brain lesion segmentation despite the abundance of unannotated data. The paper adapts a mean teacher framework with consistency training and exponential moving averaging, and reports improved ischemic stroke lesion segmentation using unannotated data.
Problem
CNN-based brain lesion segmentation requires substantial annotated data, but manual annotation is costly and time-consuming while unannotated data is often abundant.
Method
An adapted mean teacher framework trains student and teacher segmentation models with annotated segmentation loss, unannotated consistency loss, and exponential moving average teacher updates.
Results
The proposed method has the highest mean Dice coefficients among the evaluated methods, with differences from each competing method significant at p < 0.05.
Takeaways & Limitations
With 10 annotated training scans, the proposed method outperforms baseline DeepMedic trained with 20 scans and performs comparably to baseline DeepMedic trained with 30 scans.
Takeaways & Limitations
The approach generates similar input pairs using Gaussian noise, while other consistency-regularization strategies remain possible for more efficient use of unannotated data.
Abstract
from arXiv · showhide
Automated brain lesion segmentation provides valuable information for the analysis and intervention of patients. In particular, methods based on convolutional neural networks (CNNs) have achieved state-of-the-art segmentation performance. However, CNNs usually require a decent amount of annotated data, which may be costly and time-consuming to obtain. Since unannotated data is generally abundant, it is desirable to use unannotated data to improve the segmentation performance for CNNs when limited annotated data is available. In this work, we propose a semi-supervised learning (SSL) approach to brain lesion segmentation, where unannotated data is incorporated into the training of CNNs. We adapt the mean teacher model, which is originally developed for SSL-based image classification, for brain lesion segmentation. Assuming that the network should produce consistent outputs for similar inputs, a loss of segmentation consistency is designed and integrated into a self-ensembling framework. Specifically, we build a student model and a teacher model, which share the same CNN architecture for segmentation. The student and teacher models are updated alternately. At each step, the student model learns from the teacher model by minimizing the weighted sum of the segmentation loss computed from annotated data and the segmentation consistency loss between the teacher and student models computed from unannotated data. Then, the teacher model is updated by combining the updated student model with the historical information of teacher models using an exponential moving average strategy. For demonstration, the proposed approach was evaluated on ischemic stroke lesion segmentation, where it improves stroke lesion segmentation with the incorporation of unannotated data.
1 Introduction
Brain lesion segmentation benefits from CNNs but is constrained by costly, time-consuming annotation. This work adapts mean teacher semi-supervised learning to use unannotated data through segmentation consistency and self-ensembling.
- CNN-based brain lesion segmentation achieves state-of-the-art performance but requires substantial annotated data for training.Manual lesion annotation is costly and time-consuming, while unannotated data is often abundant.
- Semi-supervised learning combines limited annotated data with abundant unannotated data to improve CNN training.Prior medical-image methods enforce consistency between feature embeddings from annotated and unannotated data.
- The proposed adapted mean teacher model uses both annotated and unannotated data to improve CNN-based brain lesion segmentation.The framework is inspired by mean teacher learning for image classification and is adapted for segmentation.
- A segmentation consistency loss compares predictions for two noisy versions of the same unannotated sample, providing regularization during training.This differs from prior approaches that compare annotated and unannotated data.
- Student and teacher CNNs are updated alternately, with the student learning from annotated segmentation loss and unannotated consistency loss.The teacher records information across training steps through self-ensembling.
- The approach improves ischemic stroke lesion segmentation and outperforms competing semi-supervised segmentation strategies.The method was evaluated specifically on ischemic stroke lesion segmentation.
2 Methods
The method adapts mean teacher self-ensembling to brain lesion segmentation by combining annotated-data segmentation loss with consistency regularization from noisy unannotated data. Teacher and student CNNs share a DeepMedic architecture, while the teacher is updated by exponential moving average.
- 2.1 Backbone CNN Architecture: DeepMedic is used as the shared backbone for teacher and student models, providing dual-pathway, multi-scale three-dimensional CNN processing.It takes paired patches at normal and downsampled resolution before concatenating multi-scale features for segmentation.
- 2.2 Semi-Supervised Lesion Segmentation with an Adapted MT Model: The adapted mean teacher framework creates similar inputs by adding noise to the same unannotated sample and enforces consistent segmentation outputs.Unlike approaches comparing annotated and unannotated data, consistency is computed between two noisy versions of one unannotated sample.
- 2.2 Semi-Supervised Lesion Segmentation with an Adapted MT Model: The student is trained with a weighted sum of annotated-data cross-entropy loss and unannotated-data soft Dice consistency loss.The student receives noisy input and learns toward the teacher’s predicted probability maps, while annotated inputs use the corresponding ground-truth labels.
- 2.2 Semi-Supervised Lesion Segmentation with an Adapted MT Model: The teacher is initialized from a DeepMedic model trained on annotated data and updated after each student step using exponential moving average weights.The EMA decay is 0.99 during the ramp-up phase and 0.999 afterward, allowing rapid forgetting initially and longer memory later.
- 2.2 Semi-Supervised Lesion Segmentation with an Adapted MT Model: Consistency weighting is reduced during early training and reaches one after a ramp-up length of 400 steps because early teacher predictions may be unreliable.The adaptive coefficient limits the influence of consistency loss before the teacher generates reasonable target labels.
- 2.3 Implementation Details: Gaussian noise is applied independently at each input voxel using additive and multiplicative components to generate noisy student and teacher inputs.The noisy intensity is computed from the original intensity with additive noise ηs and multiplicative noise ηm.
3 Experiments
The experiments evaluated the adapted mean teacher approach for ischemic stroke lesion segmentation using annotated and unannotated diffusion-weighted images. Across qualitative and quantitative evaluations, the proposed method achieved better segmentation quality than competing approaches, including under different amounts of annotated training data.
- Experimental setup: The study used 246 diffusion-weighted images, including 50 manually annotated scans and 196 unannotated scans.For one experiment, 20 annotated scans were used for training, 30 for testing, and all 196 unannotated scans were included in training.
- Experimental setup: The proposed method was compared with DeepMedic, DeepMedic-UDA, and another SSL-based DeepMedic strategy.DeepMedic served as the baseline without unannotated data, while the competing strategies incorporated semi-supervised learning.
- Qualitative evaluation: The proposed method produced segmentations that better agreed with manual delineations for representative subjects with different lesion sizes.Cross-sectional results overlaid on diffusion-weighted images were compared against the gold standard and competing methods.
- Quantitative evaluation: The proposed method had the highest mean Dice coefficients on 30 test subjects, with significant differences from each competing method (p < 0.05).Means and standard deviations were computed on the test scans, and paired Student’s t-tests were used for comparisons.
- Impact of training-data amount: With 10 and 30 annotated training scans, the proposed approach achieved higher mean Dice coefficients than competing methods in all cases, with significant differences.All unannotated data were also included during training for SSL-based methods.
4 Discussion
The discussion attributes the method’s behavior to adapting mean teacher consistency learning to segmentation and using consistency between noisy versions of the same unannotated sample. The approach remained effective with fewer annotated scans, while the choice of Gaussian noise remains an open design option.
- Method adaptation: The adapted mean teacher model replaces classification consistency with a Dice-based segmentation consistency loss.The adaptation enables the mean teacher strategy to be applied to semi-supervised image segmentation.
- Robustness to limited annotations: The proposed method was robust to decreasing numbers of training scans, whereas competing adversarial SSL methods could perform worse than baseline DeepMedic with few scans.The competing methods may fail to regularize properly when the annotated scans poorly represent the desired feature and segmentation distributions.
- Robustness to limited annotations: The proposed approach was less affected by few training scans because it did not compare annotated and unannotated data directly.Its consistency assumption was applied to similar inputs, combined with self-ensembling.
- Robustness to limited annotations: With only 10 annotated scans, the proposed method outperformed baseline DeepMedic trained with 20 scans and performed comparably to baseline DeepMedic trained with 30 scans.The authors connect this result to the incorporation of unannotated data.
- Future directions: Gaussian noise generated the similar input pairs used for student and teacher models, but other consistency-enforcing strategies remain possible.The discussion specifically mentions dropout as an alternative and calls for further exploration of more efficient approaches.
5 Conclusion
The proposed semi-supervised approach combines annotated and unannotated data through alternating student–teacher updates, improving ischemic stroke lesion segmentation.
- The approach constructs teacher and student models that are updated alternately for brain lesion segmentation.
- The student minimizes annotated-data segmentation loss alongside unannotated-data consistency loss from the teacher.
- The teacher is updated using an exponential moving average strategy, and the final teacher segments test samples.
- The method improved ischemic stroke lesion segmentation by incorporating unannotated data.