Training AI Models for Automatic Segmentation and Quantification of White Matter Lesions Across Multiple Centres

Background and Objectives: White matter hyperintensities (WMH) are MRI biomarkers associated with vascular, neurodegenerative, and mixed neurological diseases. Manual WMH segmentation and volume quantification are time-consuming and operator-dependent, making artificial intelligence (AI) an attractive alternative. This study examined how WMH segmentation and volume-quantification performance changed across nested training cohorts with joint increases in participant and annotated-lesion counts. Materials and Methods: More than 18,000 WMH lesions were manually annotated by four expert raters from multicentre 3T FLAIR ADNI patient studies (n = 300: HC = 108, MCI = 108, and AD = 84) (age: 71.44 ± 7.50; education: 16.10 ± 2.37 years; 142 males). Three nnU-Net models were trained on different numbers of annotated lesions (18,034, 11,496, and 4685 lesions). Models’ performance was evaluated using the Dice Similarity Coefficient (DSC) with 5-fold cross-validation. We evaluated generalisability on an independent external dataset of n = 62 patients from European and Asian centres using quantitative and qualitative analyses via a neuroradiological 4-point visual rating scale completed by two expert raters. Results: DSC increased with training set size in both the cross-validation datasets (0.64 ± 0.03, 0.68 ± 0.04, and 0.70 ± 0.02) and the external dataset (0.67, 0.73, and 0.75), with statistically significant improvements and good-to-excellent expert-rated segmentation quality (mean score: 3.23). The model trained on 18,034 lesions achieved performance within the range reported for currently available medical devices. Conclusions: Increasing both the number of participants and annotated WMH lesions substantially improves AI-based WMH segmentation and quantification. Training on more than 18,000 lesions produced robust generalisation, expert-rated good-to-excellent segmentation quality, and performance within the range reported for currently available medical devices.

Authors

Institutions

Publication Details

Journal
Medicina
Published
2026-11-17
DOI
https://doi.org/10.3390/medicina62112253
Primary Topic
Medical Image Segmentation Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Training AI Models for Automatic Segmentation and Quantification of White Matter Lesions Across Multiple Centres

Francesca Benedetta Pizzini, Christian Salvatore, Francesco Sardanelli, Isabella Castiglioni et al.
2 links
Medicina
Medical Image Segmentation Techniques
article

Training AI Models for Automatic Segmentation and Quantification of White Matter Lesions Across Multiple Centres

Francesca Benedetta Pizzini, Christian Salvatore, Francesco Sardanelli, Isabella Castiglioni, Simona Aresta, Moreno Zanardo, Paolo Vitali, Muhammad Asim, Alberto Mosconi, Shaghayegh Adibpoor, Alzheimer’s Disease Neuroimaging Initiative
article en

Abstract

Background and Objectives: White matter hyperintensities (WMH) are MRI biomarkers associated with vascular, neurodegenerative, and mixed neurological diseases. Manual WMH segmentation and volume quantification are time-consuming and operator-dependent, making artificial intelligence (AI) an attractive alternative. This study examined how WMH segmentation and volume-quantification performance changed across nested training cohorts with joint increases in participant and annotated-lesion counts. Materials and Methods: More than 18,000 WMH lesions were manually annotated by four expert raters from multicentre 3T FLAIR ADNI patient studies (n = 300: HC = 108, MCI = 108, and AD = 84) (age: 71.44 ± 7.50; education: 16.10 ± 2.37 years; 142 males). Three nnU-Net models were trained on different numbers of annotated lesions (18,034, 11,496, and 4685 lesions). Models’ performance was evaluated using the Dice Similarity Coefficient (DSC) with 5-fold cross-validation. We evaluated generalisability on an independent external dataset of n = 62 patients from European and Asian centres using quantitative and qualitative analyses via a neuroradiological 4-point visual rating scale completed by two expert raters. Results: DSC increased with training set size in both the cross-validation datasets (0.64 ± 0.03, 0.68 ± 0.04, and 0.70 ± 0.02) and the external dataset (0.67, 0.73, and 0.75), with statistically significant improvements and good-to-excellent expert-rated segmentation quality (mean score: 3.23). The model trained on 18,034 lesions achieved performance within the range reported for currently available medical devices. Conclusions: Increasing both the number of participants and annotated WMH lesions substantially improves AI-based WMH segmentation and quantification. Training on more than 18,000 lesions produced robust generalisation, expert-rated good-to-excellent segmentation quality, and performance within the range reported for currently available medical devices.

MedicinaVol. 62(11)
University of Verona (IT), University of Milan (IT), Istituto Universitario di Studi Superiori di Pavia (IT), IRCCS Ospedale Galeazzi - Sant'Ambrogio (IT), Lega Italiana per la Lotta ai Tumori (IT), Centro Diagnostico Italiano (IT), Link Campus University (IT), University of Milano-Bicocca (IT)
Good health and well-being
Medical Image Segmentation Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon