Federative ischemic stroke segmentation as alternative to overcome domain-shift multi-institution challenges Journal: Radiology: Artificial Intelligence

Edgar Rangel, Fabio Martínez · 2025

Read Federative ischemic stroke segmentation as alternative to overcome domain-shift multi-institution challenges Journal: Radiology: Artificial Intelligence by Edgar Rangel, Fabio Martínez (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-18296.txt.

Introduction

Edgar Rangel and Fabio Martínez segment ischemic stroke lesions across hospitals without pooling the scans, in a federated learning study. Their paper from twenty twenty-five lays out the clinical stakes: stroke ranks among the top killers and disablers, and diffusion imaging guides treatment. Lesion reading still varies with demographics, scanner vendors, and annotators. Single-center models inherit one site's quirks and fail elsewhere.

Fourteen emulated centers with over two thousand studies train a shared model by averaging updates. The federated model tops centralized and rival federated rules on overlap and volume error. Performance holds across lesion types and stays reliable at unseen centers with no extra training. Domain shift, the field's nemesis, meets a workable answer.

Catalog record

CREATOREdgar Rangel, Fabio Martínez
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-18296

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Federative ischemic stroke segmentation as alternative to overcome domain-shift multi-institution challenges Journal: Radiology: Artificial Intelligence Edgar Rangel, Fabio Martínez Abstract Stroke is the second leading cause of death and the third leading cause of disability worldwide. Clinical guidelines establish diffusion resonance imaging (DWI, ADC) as the standard for localizing, characterizing, and measuring infarct volume, enabling treatment support and prognosis. Nonetheless, such lesion analysis is highly variable due to different patient demographics, scanner vendors, and expert annotations. Computational support approaches have been key to helping with the localization and segmentation of lesions. However, these strategies are dedicated solutions that learn patterns from only one institution, lacking the variability to generalize geometrical lesions shape models. Even worse, many clinical centers lack sufficient labeled samples to adjust these dedicated solutions. This work developed a collaborative framework for segmenting ischemic stroke lesions in DWI sequences by sharing knowledge from deep center-independent representations. From 14 emulated healthcare centers with 2031 studies, the FedAvg model achieved a general DSC of 0.71 ± 0.24 0.71\pm 0.24 , AVD of 5.29 ± 22.74 5.29\pm 22.74 , ALD of 2.16 ± 3.60 2.16\pm 3.60 and LF1 of 0.70 ± 0.26 0.70\pm 0.26 over all centers, outperforming both the centralized and other federated rules. Interestingly, the model demonstrated strong generalization properties, showing uniform performance across different lesion categories and reliable performance in out-of-distribution centers (with DSC of 0.64 ± 0.29 0.64\pm 0.29 and AVD of 4.44 ± 8.74 4.44\pm 8.74 without any additional training). Figure 1: The proposed approach pipeline. In (a) there are multiple heterogeneous large and limited centers with and without training capabilities, respectively. Then in (b), the large centers deliver their models to a central server which carries out a federated aggregation rule to generate a single federated model. Lastly, in (c) the federated model is delivered to large centers to fine-tune their models and repeat the process to improve the federated model, while limited centers benefit from the federated representations to generate segmentations and support experts inside the healthcare center. 1 Introduction Stroke, a disruption in the brain blood flow, is the second leading cause of death and the third leading cause of disability worldwide [ 1 ] . The ischemic stroke, caused by blood vessel occlusions, represents more than 80% of all reported stroke cases [ 2 ] . According to clinical guidelines, radiologists can discriminate between affected and healthy brain tissue from both Diffusion-Weighted Imaging sequences (DWI) and Apparent Diffusion Coefficient (ADC) parametric maps [ 3 , 4 ] . Particularly, ischemic stroke lesions are principally perceived in DWI as hyperintense signals, which are the response of intracellular cytotoxic and water accumulation. Complementary, the lesion shape and localization are confirmed by observation on complementary ADC maps [ 5 ] . In these maps, lesions appear as hypointense signals, calculated by varying the amplitude and duration of bipolar gradients of molecules ( a.k.a b-values) in DWI between 0 s/mm^{2} (DWI-B0) and 1000 s/mm^{2} (DWI-B1000) [ 6 ] . Both modalities are key for identifying ischemic infarcts within the critical thrombolytic window, especially when there exists uncertainty of onset symptoms [ 7 , 8 ] . Nonetheless, such lesion characterization is expert-dependent, which in turn is challenging due to the high variability in lesion shape, size, and location [ 5 ] . Particularly, while neuroradiologists often agree on the lesion’s core, there can be significant disagreement regarding lesions boundaries, hindering accurate volume estimation, a critical factor in treatment decisions and outcome prognosis [ 9 ] . Computational approaches have emerged as a potential alternative to mitigate lesion characterization challenges and support lesion detection over MRI studies [ 10 , 11 ] . However, these solutions learn from a dedicated set of data, typically from a unique healthcare center, with biased expert annotations [ 10 ] . Hence, a significant and variable amount of data is required in terms of patient demography, symptoms onset diversity, and even integrating annotations from several radiologists to deal with observation subjectivity [ 11 ] . A main challenge to achieve such data requirements is an effective collaboration among centers for knowledge sharing, because of limitations related to data privacy and security [ 12 ] . In fact, there exist multiple regulations (for instance, HIPAA or GDPR legislations) and standards that are constantly updated to guarantee patient privacy, but resulting each time more restrictive for center collaboration [ 13 , 14 ] . Even, each institution has proper characteristics regarding data, experts, and computational capabilities, evidencing the necessity to construct personalized or collaborative solutions that can be mapped or adapted, according to the requirements of each center [ 12 ] . Federated Learning (FL) has emerged as a promising alternative to share computational representations protecting patient privacy and enabling collaborative research across institutions by considering their own characteristics. Particularly, for stroke segmentation some seminal works have evidenced opportunities in this collaborative learning [ 15 , 16 , 17 ] . However, such approaches are limited in variability regarding expert annotations and lesions’ properties, along with the emulation of heterogeneous independent institutions given the unique source of data. In this work a federated learning framework was developed to enable collaborative knowledge from ischemic stroke segmentation models. The model showed strong generalization properties across different institutions and supported segmentation in heterogeneous lesions. The key contributions of this works are: 1. The study employed 2031 DWI and ADC neuroimages, from 14 emulated healthcare centers, based on three public and two private datasets. The institutions were divided into large and limited centers (with AI training capabilities and out-of-distribution institutions, respectively) to validate the models. 2. Federate Learning enables a collaborative platform for sharing knowledge to tackle ischemic stroke variability, leveraging heterogeneous scenarios such as patients demography, expert annotations, scanner vendors and lesions morphology. 3. The FedAvg model achieved the best performance against the centralized and other federated rules, evidencing a reliable and uniform performance in seen and out-of-distribution institutions, along with several lesion morphologies. Figure 2: Progression of the FedAvg segmentation over DWI sequences for small, medium and large volume lesions over the rounds (blue color is the expert delineation, red is from models and pink is the intersection). The last column shows segmentation from the centralized baseline. 2 Related works Diffusion-Weighted Magnetic Resonance Imaging (DWI-MRI) plays a crucial role in diagnosing ischemic stroke, evidencing intracellular cytotoxic and brain water accumulation as hyperintense signals in the images proportional to stroke advanced stages [ 8 ] . However, DWI sequences may report artifacts such as T2-shine-through or T2-blackout voxels, misleading the localization of ischemic lesions [ 18 ] . To complement such analysis, the Apparent Diffusion Coefficient (ADC) parametric map avoids these artifacts on DWI, showing the ischemic lesion as hypointense signals and allowing to confirm the localization and shape of flow obstruction, observed from DWI-B0 and DWI-B1000 sequences [ 4 ] . Findings observed in DWI are manually delineated and confirmed regarding the ADC map for quantifying the lesion volume, to support diagnosis and treatment planning, especially when symptom onset is unclear [ 7 , 8 ] . The lesion segmentation support can be obtained by using classic machine learning estimators, such as DecisionTree and RandomForest, or even by following computer vision strategies such as graph-cuts

Read the complete text