Automated Detection and Quantification of Hemorrhagic Transformation After Endovascular Thrombectomy

BACKGROUND: Hemorrhagic transformation after endovascular thrombectomy is a principal determinant of clinical outcome. Although artificial intelligence (AI) algorithms for spontaneous hemorrhage detection have demonstrated high diagnostic accuracy, none have been rigorously validated for postprocedural hemorrhagic transformation across diverse imaging modalities. METHODS: We conducted a multicenter diagnostic accuracy study within a nationwide stroke registry (18 centers, June 2022–December 2023). Patients who underwent endovascular thrombectomy for anterior circulation large-vessel occlusion and received follow-up noncontrast computed tomography, gradient-recalled echo, or susceptibility-weighted imaging within 168 hours were included. AI-derived hemorrhage volumes (mL) from a regulatory-cleared algorithm (noncontrast computed tomography) and newly developed algorithms (gradient-recalled echo, susceptibility-weighted imaging) were compared against expert-determined ECASS (European Cooperative Acute Stroke Study) classification. Diagnostic performance was assessed by sensitivity, specificity, and receiver operating characteristic analysis. Three-month modified Rankin Scale scores were evaluated for volume–outcome association. RESULTS: Among 1490 patients (median age 73 years; 57.4% male), hemorrhagic transformation was present in 41.4% and parenchymal hemorrhage in 11.1%. Parenchymal hemorrhage detection sensitivity exceeded 94% across all modalities (noncontrast computed tomography, 95.4%; gradient-recalled echo, 94.4%; and susceptibility-weighted imaging, 98.3%), with area under the curves of 0.900, 0.943, and 0.953, respectively. Youden-optimized volume thresholds for parenchymal hemorrhage ranged from 1.6 to 1.8 mL. AI-derived volume demonstrated an exposure-response association with 3-month modified Rankin Scale (Spearman ρ=0.353; P <0.001); good outcome (modified Rankin Scale score, 0–2) declined from 61.8% to 6.7% across increasing volume categories. Among ECASS 0 cases, AI-positive patients had significantly worse outcomes than true-negatives (good outcome 48.2% versus 67.2%; P <0.001). CONCLUSIONS: AI-based hemorrhage quantification demonstrated robust sensitivity for clinically significant parenchymal hemorrhage after endovascular thrombectomy across noncontrast computed tomography, gradient-recalled echo, and susceptibility-weighted imaging, with high diagnostic accuracy across all 3 modalities. AI-derived hemorrhage volume exhibited an exposure-response relationship with functional outcome and identified a prognostically distinct subgroup among expert-classified no hemorrhagic transformation cases, supporting its potential as a continuous prognostic biomarker.

Authors

Institutions

Publication Details

Journal
Stroke Vascular and Interventional Neurology
Published
2026-11-19
DOI
https://doi.org/10.1161/svin.126.002506
Primary Topic
Acute Ischemic Stroke Management
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Automated Detection and Quantification of Hemorrhagic Transformation After Endovascular Thrombectomy

Wook‐Joo Kim, Hae-Bong Jeong, Kyung‐Ho Yu, Dong‐Seok Gwak et al.
3 links
Stroke Vascular and Interventional Neurology
Acute Ischemic Stroke Management
article

Automated Detection and Quantification of Hemorrhagic Transformation After Endovascular Thrombectomy

Wook‐Joo Kim, Hae-Bong Jeong, Kyung‐Ho Yu, Dong‐Seok Gwak, Jihoon Kang, Mi Sun Oh, Wi‐Sun Ryu, Tai Hwan Park, Kwang‐Yeol Park, Keun‐Sik Hong, Soo Joo Lee, Jeong‐Ho Hong, Leonard Sunwoo, Joong‐Goo Kim, Sung‐Il Sohn, Dong‐Eog Kim, Jae‐Kwan Cha, Dong‐Ick Shin, Kyu Sun Yum, Hee‐Joon Bae, Jay Chol Choi, Doo Hyuk Kwon, Jun Yup Kim, Hong‐Kyun Park, Kyusik Kang, Keon‐Joo Lee, Minwoo Lee, Yong‐Jin Cho, Beom Joon Kim, Jeong‐Yoon Lee, Hee-Jung Ha, Kyungbok Lee, Jee Hyun Kwon, Chi Kyung Kim, Myung Jae Lee, Jun Lee, Chulho Kim, Hyunsoo Kim, Jae Guk Kim, Sang-Hwa Lee
article en

Abstract

BACKGROUND: Hemorrhagic transformation after endovascular thrombectomy is a principal determinant of clinical outcome. Although artificial intelligence (AI) algorithms for spontaneous hemorrhage detection have demonstrated high diagnostic accuracy, none have been rigorously validated for postprocedural hemorrhagic transformation across diverse imaging modalities. METHODS: We conducted a multicenter diagnostic accuracy study within a nationwide stroke registry (18 centers, June 2022–December 2023). Patients who underwent endovascular thrombectomy for anterior circulation large-vessel occlusion and received follow-up noncontrast computed tomography, gradient-recalled echo, or susceptibility-weighted imaging within 168 hours were included. AI-derived hemorrhage volumes (mL) from a regulatory-cleared algorithm (noncontrast computed tomography) and newly developed algorithms (gradient-recalled echo, susceptibility-weighted imaging) were compared against expert-determined ECASS (European Cooperative Acute Stroke Study) classification. Diagnostic performance was assessed by sensitivity, specificity, and receiver operating characteristic analysis. Three-month modified Rankin Scale scores were evaluated for volume–outcome association. RESULTS: Among 1490 patients (median age 73 years; 57.4% male), hemorrhagic transformation was present in 41.4% and parenchymal hemorrhage in 11.1%. Parenchymal hemorrhage detection sensitivity exceeded 94% across all modalities (noncontrast computed tomography, 95.4%; gradient-recalled echo, 94.4%; and susceptibility-weighted imaging, 98.3%), with area under the curves of 0.900, 0.943, and 0.953, respectively. Youden-optimized volume thresholds for parenchymal hemorrhage ranged from 1.6 to 1.8 mL. AI-derived volume demonstrated an exposure-response association with 3-month modified Rankin Scale (Spearman ρ=0.353; P <0.001); good outcome (modified Rankin Scale score, 0–2) declined from 61.8% to 6.7% across increasing volume categories. Among ECASS 0 cases, AI-positive patients had significantly worse outcomes than true-negatives (good outcome 48.2% versus 67.2%; P <0.001). CONCLUSIONS: AI-based hemorrhage quantification demonstrated robust sensitivity for clinically significant parenchymal hemorrhage after endovascular thrombectomy across noncontrast computed tomography, gradient-recalled echo, and susceptibility-weighted imaging, with high diagnostic accuracy across all 3 modalities. AI-derived hemorrhage volume exhibited an exposure-response relationship with functional outcome and identified a prognostically distinct subgroup among expert-classified no hemorrhagic transformation cases, supporting its potential as a continuous prognostic biomarker.

Stroke Vascular and Interventional Neurology
Chonnam National University (KR), Ewha Womans University (KR), Seoul National University (KR), Chungbuk National University (KR), Eulji University (KR), Soonchunhyang University (KR), Dong-A University Hospital (KR), Soonchunhyang University Hospital Seoul (KR), New Generation University College (ET), Seoul National University Bundang Hospital (KR), Sacred Heart Hospital (NG), Seoul Medical Center (KR), Chonnam National University Hospital (KR), Dongguk University Ilsan Hospital (KR), Inje University Ilsan Paik Hospital (KR), Daejeon Eulji Medical Center, Eulji University (KR), Ulsan University Hospital (KR), Yeungnam University Medical Center (KR), Chung-Ang University Hospital (KR), Hallym University Sacred Heart Hospital (KR), Chungbuk National University Hospital (KR), Hallym University Chuncheon Sacred Heart Hospital (KR), Jeju National University Hospital (KR), Eulji University Nowon Eulji Medical Center (KR), Keimyung University Dongsan Hospital (KR), Ewha Womans University Seoul Hospital (KR), Keimyung University (KR), Chung-Ang University (KR)
Good health and well-being
Acute Ischemic Stroke Management
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