10 January 2022 in IEEE Transactions on Multimedia
Recent advances in unsupervised domain adaptation (UDA) techniques have witnessed great success in cross-domain computer vision tasks, enhancing the generalization ability of data-driven deep learning architectures by bridging the domain distribution gaps. For the UDA-based cross-domain object detection methods, the majority of them alleviate the domain bias by inducing the domain-invariant feature generation via adversarial learning strategy. However, their domain discriminators have limited classification ability due to the unstable adversarial training process. Therefore, the extracted features induced by them cannot be perfectly domain-invariant and still contain domain-private factors, bringing obstacles to further alleviate the cross-domain discrepancy. To tackle this issue, we design a Domain Disentanglement Faster-RCNN (DDF) to eliminate the source-specific information in the features for detection task learning. Our DDF method facilitates the feature disentanglement at the global and local stages, with a Global Triplet Disentanglement (GTD) module and an Instance Similarity Disentanglement (ISD) module, respectively. By outperforming state-of-the-art methods on four benchmark UDA object detection tasks, our DDF method is demonstrated to be effective with wide applicability.
1 March 2019 in Ophthalmology
19 October 2023 in npj Digital Medicine
Modern management of MS targets No Evidence of Disease Activity (NEDA): no clinical relapses, no magnetic resonance imaging (MRI) disease activity and no disability worsening. While MRI is the principal tool available to neurologists for monitoring clinically silent MS disease activity and, where appropriate, escalating treatment, standard radiology reports are qualitative and may be insensitive to the development of new or enlarging lesions. Existing quantitative neuroimaging tools lack adequate clinical validation. In 397 multi-center MRI scan pairs acquired in routine practice, we demonstrate superior case-level sensitivity of a clinically integrated AI-based tool over standard radiology reports (93.3% vs 58.3%), relative to a consensus ground truth, with minimal loss of specificity. We also demonstrate equivalence of the AI-tool with a core clinical trial imaging lab for lesion activity and quantitative brain volumetric measures, including percentage brain volume loss (PBVC), an accepted biomarker of neurodegeneration in MS (mean PBVC −0.32% vs −0.36%, respectively), whereas even severe atrophy (>0.8% loss) was not appreciated in radiology reports. Finally, the AI-tool additionally embeds a clinically meaningful, experiential comparator that returns a relevant MS patient centile for lesion burden, revealing, in our cohort, inconsistencies in qualitative descriptors used in radiology reports. AI-based image quantitation enhances the accuracy of, and value-adds to, qualitative radiology reporting. Scaled deployment of these tools will open a path to precision management for patients with MS.
12 April 2016 in Journal of Neurology, Neurosurgery & Psychiatry
Background Whole brain volume (WBV) estimates in patients with multiple sclerosis (MS) correlate more robustly with clinical disability than traditional, lesion-based metrics. Numerous algorithms to measure WBV have been developed over the past two decades. We compare Structural Image Evaluation using Normalisation of Atrophy-Cross-sectional (SIENAX) to NeuroQuant and MSmetrix, for assessment of cross-sectional WBV in patients with MS. Methods MRIs from 61 patients with relapsing-remitting MS and 2 patients with clinically isolated syndrome were analysed. WBV measurements were calculated using SIENAX, NeuroQuant and MSmetrix. Statistical agreement between the methods was evaluated using linear regression and Bland-Altman plots. Precision and accuracy of WBV measurement was calculated for (1) NeuroQuant versus SIENAX and (2) MSmetrix versus SIENAX. Results Precision (Pearson's r) of WBV estimation for NeuroQuant and MSmetrix versus SIENAX was 0.983 and 0.992, respectively. Accuracy (Cb) was 0.871 and 0.994, respectively. NeuroQuant and MSmetrix showed a 5.5% and 1.0% volume difference compared with SIENAX, respectively, that was consistent across low and high values. Conclusions In the analysed population, NeuroQuant and MSmetrix both quantified cross-sectional WBV with comparable statistical agreement to SIENAX, a well-validated cross-sectional tool that has been used extensively in MS clinical studies.
6 April 2022 in Medical Image Analysis
17 April 2024 in Artificial Intelligence in Medicine
Accurately measuring the evolution of Multiple Sclerosis (MS) with magnetic resonance imaging (MRI) critically informs understanding of disease progression and helps to direct therapeutic strategy. Deep learning models have shown promise for automatically segmenting MS lesions, but the scarcity of accurately annotated data hinders progress in this area. Obtaining sufficient data from a single clinical site is challenging and does not address the heterogeneous need for model robustness. Conversely, the collection of data from multiple sites introduces data privacy concerns and potential label noise due to varying annotation standards. To address this dilemma, we explore the use of the federated learning framework while considering label noise. Our approach enables collaboration among multiple clinical sites without compromising data privacy under a federated learning paradigm that incorporates a noise-robust training strategy based on label correction. Specifically, we introduce a Decoupled Hard Label Correction (DHLC) strategy that considers the imbalanced distribution and fuzzy boundaries of MS lesions, enabling the correction of false annotations based on prediction confidence. We also introduce a Centrally Enhanced Label Correction (CELC) strategy, which leverages the aggregated central model as a correction teacher for all sites, enhancing the reliability of the correction process. Extensive experiments conducted on two multi-site datasets demonstrate the effectiveness and robustness of our proposed methods, indicating their potential for clinical applications in multi-site collaborations to train better deep learning models with lower cost in data collection and annotation.