Transfer Learning for Domain Adaptation in MRI: Application in Brain Lesion Segmentation (1702.07841v1)

Abstract: Magnetic Resonance Imaging (MRI) is widely used in routine clinical diagnosis and treatment. However, variations in MRI acquisition protocols result in different appearances of normal and diseased tissue in the images. Convolutional neural networks (CNNs), which have shown to be successful in many medical image analysis tasks, are typically sensitive to the variations in imaging protocols. Therefore, in many cases, networks trained on data acquired with one MRI protocol, do not perform satisfactorily on data acquired with different protocols. This limits the use of models trained with large annotated legacy datasets on a new dataset with a different domain which is often a recurring situation in clinical settings. In this study, we aim to answer the following central questions regarding domain adaptation in medical image analysis: Given a fitted legacy model, 1) How much data from the new domain is required for a decent adaptation of the original network?; and, 2) What portion of the pre-trained model parameters should be retrained given a certain number of the new domain training samples? To address these questions, we conducted extensive experiments in white matter hyperintensity segmentation task. We trained a CNN on legacy MR images of brain and evaluated the performance of the domain-adapted network on the same task with images from a different domain. We then compared the performance of the model to the surrogate scenarios where either the same trained network is used or a new network is trained from scratch on the new dataset.The domain-adapted network tuned only by two training examples achieved a Dice score of 0.63 substantially outperforming a similar network trained on the same set of examples from scratch.

An Overview of Transfer Learning for Domain Adaptation in MRI: Applications in Brain Lesion Segmentation

The paper "Transfer Learning for Domain Adaptation in MRI: Application in Brain Lesion Segmentation" by Ghafoorian et al. addresses the challenges of deploying convolutional neural networks (CNNs) in medical imaging where data distribution shifts, caused by variations in MRI acquisition protocols, are prevalent. These distributional shifts can impede the performance of CNNs when applied to datasets outside their original domain, thus necessitating effective domain adaptation techniques. This paper investigates transfer learning (TL) as a strategy to enable CNNs, initially trained on one MRI protocol, to maintain efficacy across diverse imaging protocols.

Key Investigations and Methodology

The central inquiry of this research revolves around two questions: the minimal new domain data necessary for effective adaptation and the extent of model re-training needed with available new domain samples. To explore these concerns, the authors conducted experiments in white matter hyperintensity (WMH) segmentation, utilizing baseline and follow-up MRI datasets with differing acquisition parameters.

The paper employs a CNN model trained on a source domain (baseline MRI data) and evaluates its adaptation to a target domain (follow-up MRI data). Their approach is rigorously tested against two baseline scenarios: direct application of the source-trained model to the target data and training a new model from scratch on the target data. Additionally, they varied training set sizes in the target domain to examine performance impacts on domain-adapted models.

Results and Analysis

The results indicate that transferring knowledge from a source domain model, even when fine-tuned with minimal target domain data, outperforms a model trained from scratch under equivalent conditions. Remarkably, with only two target domain training images, the domain-adapted model achieved a Dice score of 0.63, in contrast to a mere 0.15 when trained from scratch. The authors also discern that while domain adaptation generally benefits from tuning deeper network layers, the adaptation is most successful with limited additional data when focusing on retraining only the network's tail (dense layers). As the availability of target domain data increases, retraining deeper into the earlier convolutional layers becomes more beneficial.

Implications and Future Directions

The findings demonstrate that transfer learning can mitigate the limitations posed by domain shifts in medical imaging, enhancing the feasibility of utilizing legacy datasets across evolving MRI protocols. This has practical implications for clinical settings where resource and data constraints are prevalent. Theory-wise, it advances our understanding of representation transfer between domains with CNN architectures, especially concerning the layers to be fine-tuned contingent upon the dataset size.

Looking forward, this work prompts the investigation of more nuanced domain adaptation techniques, perhaps involving adversarial networks or advanced generative models to further refine cross-domain generalizability. Furthermore, strategies that combine TL with other methods like few-shot learning could prove beneficial in further minimizing the requirement for extensive new domain data.

Overall, the research provides a valuable contribution to medical image analysis, specifically in leveraging deep learning's potential in situations where direct application of existing models falls short due to domain shift challenges.