Source-linked AI summary
One-shot domain adaptation in multiple sclerosis lesion segmentation using convolutional neural networks
Sergi Valverde, Mostafa Salem, Mariano Cabezas, Deborah Pareto, Joan C. Vilanova, Lluís Ramió-Torrentà, Àlex Rovira, Joaquim Salvi, Arnau Oliver, Xavier Lladó
TL;DR
CNN-based MS lesion segmentation can lose accuracy on unseen MRI domains, creating a need for adaptation with limited target-domain annotation. The paper transfers a source CNN trained on two public MS datasets and retrains selected fully connected layers on new domains. On ISBI2015, one-shot adaptation achieved accuracy comparable to a human expert and to CNNs trained on the full available training set.
Problem
CNN lesion-segmentation accuracy decreases on MRI domains differing in scanners, protocols, contrast, noise, or resolution, while new-domain annotation is time-consuming.
Method
A cascaded CNN source model trained on two public MS datasets is adapted to unseen domains by retraining selected fully connected layers using target-domain images and labels.
Results
On ISBI2015, the one-shot domain-adapted model had accuracy comparable to a human expert rater and to CNN methods trained on a wide set of training data.
Takeaways & Limitations
The approach supports MS lesion segmentation in different clinical settings with reduced amounts of annotated data and may reduce manual-labeling time and economic costs.
Takeaways & Limitations
For datasets with at least 3 ml of lesion voxels, the best results used the last two or all FC layers, whereas datasets below 3 ml favored retraining only the last layer to reduce over-fitting.
Abstract
from arXiv · showhide
In recent years, several convolutional neural network (CNN) methods have been proposed for the automated white matter lesion segmentation of multiple sclerosis (MS) patient images, due to their superior performance compared with those of other state-of-the-art methods. However, the accuracies of CNN methods tend to decrease significantly when evaluated on different image domains compared with those used for training, which demonstrates the lack of adaptability of CNNs to unseen imaging data. In this study, we analyzed the effect of intensity domain adaptation on our recently proposed CNN-based MS lesion segmentation method. Given a source model trained on two public MS datasets, we investigated the transferability of the CNN model when applied to other MRI scanners and protocols, evaluating the minimum number of annotated images needed from the new domain and the minimum number of layers needed to re-train to obtain comparable accuracy. Our analysis comprised MS patient data from both a clinical center and the public ISBI2015 challenge database, which permitted us to compare the domain adaptation capability of our model to that of other state-of-the-art methods. For the ISBI2015 challenge, our one-shot domain adaptation model trained using only a single image showed a performance similar to that of other CNN methods that were fully trained using the entire available training set, yielding a comparable human expert rater performance. We believe that our experiments will encourage the MS community to incorporate its use in different clinical settings with reduced amounts of annotated data. This approach could be meaningful not only in terms of the accuracy in delineating MS lesions but also in the related reductions in time and economic costs derived from manual lesion labeling.
1. Introduction
CNNs have achieved strong performance for automated MS lesion segmentation, but their accuracy often drops on unseen MRI domains. This study evaluates supervised adaptation across unrelated MS databases and the amount of target-domain annotation and retraining required.
- MRI lesion inspection is time-consuming, error-prone, and variable across expert raters, motivating automated segmentation methods.
- CNNs avoid manual feature engineering and have achieved top-ranking performance in international MS lesion challenges.
- CNN accuracy often fails to generalize across scanners, acquisition protocols, contrast, noise levels, and resolutions.
- Prior domain-adaptation studies did not focus on completely unrelated MS image domains or MS lesion-segmentation tasks.
- The study trains a source CNN on two public MS databases and evaluates transfer to unseen clinical and ISBI2015 domains while varying annotated images and retrained layers.
2. Materials and methods
The method uses a cascaded CNN trained on multi-sequence 3D patches, then adapts selected fully connected layers to target MRI domains. Reusing convolutional representations reduces target-domain parameters and supports training with fewer annotated images.
- CNN architecture: The cascaded framework trains one CNN to detect candidate lesion voxels and a second CNN to reduce false positives.
- CNN architecture: The modified architecture doubles convolutional-stack depth and adds fully connected layers of sizes 128 and 64, producing approximately 470K parameters.
- CNN architecture: Training uses balanced multi-channel FLAIR and T1-w 3D patches of size 11 × 11 × 11, followed by a second model trained on misclassified lesion voxels.
- Source model: The source model is fully trained on 35 images from the MICCAI2008 and MICCAI2016 public MS lesion datasets.
- Supervised domain adaptation: Domain adaptation retrains the last FC layer, the last two FC layers, or all FC layers using target-domain images and labels.
- Supervised domain adaptation: Reusing convolutional features lowers the number of target-domain weights to optimize, enabling reduced-image training without over-fitting the model.
3.1. Clinical MS dataset
The clinical MS experiment tested one-shot and partial domain adaptation by varying annotated training images and retrained fully connected layers. With more target-domain data, models retraining the last two or all fully connected layers approached the fully trained CNN, while sensitivity gains increased false positives.
- Experimental setup: The clinical dataset was split into 30 training and 30 testing images, and target models were evaluated against lesion annotations using DSC, sensitivity, and precision.Comparisons included the unretrained source model, fully trained target CNN, LST, and SLS.
- One-shot adaptation: One-shot adaptation retrained each configuration using a single image with lesion burdens ranging from 0.5 to 18 ml.The study compared models retraining 1, 2, or all fully connected layers.
- Training-data effect: Target-domain training-image counts ranged from 1 to 30, allowing the effect of additional annotated data on adaptation accuracy to be assessed.The reported configurations used 1, 2, 5, 10, 15, or all available training images.
- Results: With more target-domain data, models retraining the last two or all fully connected layers performed similarly to the fully trained CNN pipeline.Figure 3 summarizes mean DSC, sensitivity, and precision over 30 testing images.
- Results: Retrained models were generally more sensitive to white-matter lesions, but this came with more false-positive outcomes.This trade-off was observed in the sensitivity and precision results.
3.2. ISBI 2015 dataset
The ISBI2015 evaluation tested one-shot adaptation by retraining selected fully connected layers with one image from the target domain. The best adapted model approached fully trained CNN performance and human-level challenge results.
- Experimental design: One-shot adaptation retrained 1, 2, or all fully connected layers using one image from each of five ISBI training subjects.The five images had lesion volumes ranging from 2.3 to 26.8 ml.
- Domain transfer: The source model without target-domain retraining showed poor accuracy on the unseen ISBI domain.Performance was compared against models retrained with different layer configurations and the fully trained reference model.
- Adaptation results: The best performance against the silver masks was obtained by retraining all fully connected layers.Figure 4 additionally illustrates predictions from models retrained only on the last fully connected layer.
- Challenge comparison: The one-shot model achieved an overall score of 90.3, similar to fully trained CNN methods and comparable to human performance.The challenge considered overall scores above 90 comparable to human performance.
4. Discussion
The discussion attributes successful adaptation to reusable convolutional features and shows that performance improves with more target-domain lesion samples. Even a single image produced high performance, while layer selection depended on lesion volume.
- Interpretation: Reusing convolutional features reduced the number of target-domain parameters and allowed fitting adapted networks with few samples without over-fitting.The source architecture contained approximately 470K parameters, but only selected fully connected layers needed retraining.
- Effect of training data: Increasing lesion samples improved DSC, sensitivity, and precision across the adapted models.Additional lesion and normal-appearing patches also reduced false-positive lesion voxels.
- Effect of training data: Adapting two or all fully connected layers became more effective as lesion samples increased, whereas one-layer adaptation remained effective with very small datasets.For datasets containing at least 3000 lesion voxels, the last two or all fully connected layers performed best; below 3 ml, retraining only the last layer reduced over-fitting.
- One-shot performance: A single-image adapted model achieved accuracy similar to other fully trained CNN methods and comparable to human performance in the ISBI2015 challenge.The result extended the source model into the target domain using a reduced training dataset.
- Scope and practice: The fixed source-model hyperparameters were not fine-tuned for the adapted models.The authors present the observed layer-selection behavior as a rule of thumb for future settings.
5. Conclusions
The study evaluated transfer from two public MS datasets to private and ISBI2015 domains while varying annotated-image and retraining-layer requirements. One target-domain image produced human-comparable ISBI2015 performance similar to CNNs trained on much larger datasets.
- Study aim: The study evaluated how many annotated target-domain images and retrained layers were needed for comparable accuracy.The source model was trained on two public MS datasets and tested on private and ISBI2015 data.
- Main finding: The proposed model transferred knowledge to new image domains even when only a single target-domain training image was available.This conclusion covered the study’s domain-adaptation experiments.
- Main finding: On ISBI2015, one-shot adaptation achieved human-comparable accuracy similar to CNN methods trained with substantially more training data.The authors suggest applicability in clinical settings with reduced annotated data and potential reductions in manual-labeling time and costs.