[Paper Review] Domain Adaptation using Silver Standard Masks for Lateral Ventricle Segmentation in FLAIR MRI

Owen Crystal, Pejman Maralani+3 more|arXiv (Cornell University)|Jul 17, 2023

Cardiac Imaging and DiagnosticsMedicine3 citations

TL;DR

This paper proposes a novel domain adaptation method for lateral ventricle segmentation in FLAIR MRI using silver standard (SS) masks generated via a conventional image processing algorithm to improve transfer learning performance across multi-centre datasets. By pre-training on SS masks from the target domain and fine-tuning on gold standard (GS) data from a source domain (CAIN), the SS+GS model achieved a mean DSC of 0.89 with 4% higher accuracy than GS-only models, demonstrating robustness and consistency across diverse clinical sites.

ABSTRACT

Lateral ventricular volume (LVV) is an important biomarker for clinical investigation. We present the first transfer learning-based LVV segmentation method for fluid-attenuated inversion recovery (FLAIR) MRI. To mitigate covariate shifts between source and target domains, this work proposes an domain adaptation method that optimizes performance on three target datasets. Silver standard (SS) masks were generated from the target domain using a novel conventional image processing ventricular segmentation algorithm and used to supplement the gold standard (GS) data from the source domain, Canadian Atherosclerosis Imaging Network (CAIN). Four models were tested on held-out test sets from four datasets: 1) SS+GS: trained on target SS masks and fine-tuned on source GS masks, 2) GS+SS: trained on source GS masks and fine-tuned on target SS masks, 3) trained on source GS (GS CAIN Only) and 4) trained on target SS masks (SS Only). The SS+GS model had the best and most consistent performance (mean DSC = 0.89, CoV = 0.05) and showed significantly (p < 0.05) higher DSC compared to the GS-only model on three target domains. Results suggest pre-training with noisy labels from the target domain allows the model to adapt to the dataset-specific characteristics and provides robust parameter initialization while fine-tuning with GS masks allows the model to learn detailed features. This method has wide application to other medical imaging problems where labeled data is scarce, and can be used as a per-dataset calibration method to accelerate wide-scale adoption.

Motivation & Objective

• To address domain shift in lateral ventricle segmentation across multi-centre FLAIR MRI datasets due to variations in scanning protocols and hardware.
• To overcome the scarcity of gold standard (GS) annotations in target domains by generating reliable silver standard (SS) masks using a novel image processing algorithm.
• To develop the first FLAIR-only deep learning-based LVV segmentation model using transfer learning for improved generalization.
• To enable per-site calibration of AI models using unlabeled target data, accelerating wide-scale clinical deployment.

Proposed method

• A novel conventional image processing-based (IPB) algorithm was developed to automatically generate silver standard (SS) masks for lateral ventricle segmentation in unlabeled FLAIR MRI volumes.
• SS masks were generated from target domain datasets (ADNI, CCNA, ONDRI) to supplement gold standard (GS) data from the source domain (CAIN).
• Four models were trained and evaluated: (1) SS+GS (pre-train on SS, fine-tune on GS), (2) GS+SS (pre-train on GS, fine-tune on SS), (3) GS CAIN Only, and (4) SS Only.
• Transfer learning was applied by pre-training on SS masks from the target domain to adapt to site-specific image characteristics before fine-tuning on high-quality GS annotations.
• The SS+GS model was selected as the optimal configuration, leveraging noisy SS labels for robust initialization and GS data for detailed feature learning.
• Performance was evaluated using Dice Similarity Coefficient (DSC), mean volume difference, and coefficient of variation (CoV) across four multi-centre datasets.

Experimental results

Research questions

• RQ1 Can silver standard masks generated from unlabeled target domain data improve domain adaptation in lateral ventricle segmentation on FLAIR MRI?
• RQ2 Does pre-training on SS masks from the target domain followed by fine-tuning on GS data from a source domain yield better performance than standard transfer learning with GS-only pre-training?
• RQ3 How does the proposed method compare to GS-only models in terms of DSC, consistency, and robustness across multi-centre datasets?
• RQ4 To what extent do SS masks help mitigate covariate shift caused by differences in scanner hardware, protocols, and image intensity distributions?
• RQ5 Can this method be generalized as a per-site calibration framework for other medical imaging AI applications with limited labeled data?

Key findings

• The SS+GS model achieved the highest and most consistent performance with a mean DSC of 0.89 and a coefficient of variation (CoV) of 0.05 across three target datasets.
• The SS+GS model showed a statistically significant (p < 0.05) 4% improvement in DSC compared to the GS-only model on all three target domains.
• The SS+GS model outperformed other configurations in segmenting small and morphologically complex ventricular regions, particularly in inferior and superior slices.
• The model demonstrated robustness with the smallest mean volume difference and lowest deviation across datasets, indicating high reliability and consistency.
• The IPB algorithm generated SS masks rapidly (5–7 minutes for 100 volumes), offering a scalable alternative to manual annotation, though it may fail in cases of significant midline shift.
• The study confirms that pre-training on noisy, domain-specific SS masks improves model generalization and parameter initialization, outperforming ImageNet-style pre-training for medical imaging tasks.