[Paper Review] A Self-Adaptive Network For Multiple Sclerosis Lesion Segmentation From Multi-Contrast MRI With Various Imaging Protocols

Yushan Feng, Huitong Pan+2 more|arXiv (Cornell University)|Nov 19, 2018

Spectroscopy Techniques in Biomedical and Chemical Research5 references4 citations

TL;DR

This paper proposes a self-adaptive deep neural network for multiple sclerosis lesion segmentation from multi-contrast MRI that dynamically handles missing or variable MRI sequences during deployment. Using a novel sequence dropout technique, the model maintains high performance across diverse imaging protocols, achieving a Dice score of 0.684 on the ISBI 2015 challenge and demonstrating FLAIR as the most critical sequence for segmentation accuracy.

ABSTRACT

Deep neural networks (DNN) have shown promises in the lesion segmentation of multiple sclerosis (MS) from multicontrast MRI including T1, T2, proton density (PD) and FLAIR sequences. However, one challenge in deploying such networks into clinical practice is the variability of imaging protocols, which often differ from the training dataset as certain MRI sequences may be unavailable or unusable. Therefore, trained networks need to adapt to practical situations when imaging protocols are different in deployment. In this paper, we propose a DNN-based MS lesion segmentation framework with a novel technique called sequence dropout which can adapt to various combinations of input MRI sequences during deployment and achieve the maximal possible performance from the given input. In addition, with this framework, we studied the quantitative impact of each MRI sequence on the MS lesion segmentation task without training separate networks. Experiments were performed using the IEEE ISBI 2015 Longitudinal MS Lesion Challenge dataset and our method is currently ranked 2nd with a Dice similarity coefficient of 0.684. Furthermore, we showed our network achieved the maximal possible performance when one sequence is unavailable during deployment by comparing with separate networks trained on the corresponding input MRI sequences. In particular, we discovered T1 and PD have minor impact on segmentation performance while FLAIR is the predominant sequence. Experiments with multiple missing sequences were also performed and showed the robustness of our network.

Motivation & Objective

• Address the challenge of variable MRI imaging protocols in clinical deployment of deep learning models for MS lesion segmentation.
• Develop a deep learning framework that adapts to missing or inconsistent MRI sequences without retraining.
• Quantify the contribution of each MRI sequence (T1, T2, PD, FLAIR) to segmentation performance without training separate models.
• Ensure robust performance across diverse clinical scenarios where certain sequences may be unavailable or unusable.

Proposed method

• Introduce a sequence dropout technique that simulates missing MRI sequences during inference, enabling the network to generalize across variable input combinations.
• Design a self-adaptive network architecture that learns to prioritize informative sequences and compensate for absent ones.
• Train the model end-to-end on multi-contrast MRI data (T1, T2, PD, FLAIR) using the ISBI 2015 dataset to ensure robustness.
• Use a unified network to evaluate performance across all possible input sequence combinations, avoiding the need for separate models per configuration.
• Apply gradient-based analysis to assess the relative importance of each MRI sequence on segmentation output.
• Leverage attention mechanisms implicitly through network architecture to adaptively weight input sequences based on availability and relevance.

Experimental results

Research questions

• RQ1 How does the proposed self-adaptive network perform when certain MRI sequences are missing during inference?
• RQ2 Which MRI sequence (T1, T2, PD, FLAIR) contributes most significantly to accurate MS lesion segmentation?
• RQ3 Can a single unified network achieve near-optimal performance across all possible combinations of input MRI sequences?
• RQ4 How does the performance degrade when multiple sequences are missing, and can the model still maintain robustness?
• RQ5 To what extent can the network adapt to clinical variability in imaging protocols without retraining?

Key findings

• The proposed method achieved a Dice similarity coefficient of 0.684 on the ISBI 2015 Longitudinal MS Lesion Challenge, ranking 2nd among submitted methods.
• When FLAIR was available, the model achieved the highest performance, indicating its dominant role in lesion detection.
• T1 and PD sequences contributed minimally to segmentation accuracy, with their absence having little impact on overall performance.
• The model maintained high performance even when one or more sequences were missing, outperforming separate models trained on partial input sets.
• Experiments with multiple missing sequences confirmed the network’s robustness and adaptability to real-world clinical variability.
• The sequence dropout technique enabled the model to generalize effectively across diverse imaging protocols without retraining.