A Fully Automated Penumbra Segmentation Tool

Kartheeban Nagenthiraja; Lars Riisgaard Ribe; Kristina Dupont Hougaard; Leif Østergaard; Kim Mouridsen

A Fully Automated Penumbra Segmentation Tool

Kartheeban Nagenthiraja, Lars Riisgaard Ribe, Kristina Dupont Hougaard, Leif Østergaard, Kim Mouridsen

Publikation: Konferencebidrag › Konferenceabstrakt til konference › Forskning › peer review

Abstract

Introduction: Perfusion- and diffusion weighted MRI (PWI/DWI) is widely used to select patients who are likely to benefit from recanalization therapy. The visual identification of PWI-DWI-mismatch tissue depends strongly on the observer, prompting a need for software, which estimates potentially salavageable tissue, quickly and accurately. We present a fully Automated Penumbra Segmentation (APS) algorithm using PWI and DWI images. We compare automatically generated PWI-DWI mismatch mask to mask outlined manually by experts, in 168 patients. Method: The algorithm initially identifies PWI lesions by applying a connected component-labeling algorithm on thresholded Time-To-Peak maps (TTP > 4 seconds). Smooth PWI lesion borders are then identified by applying a level-set algorithm on the initial PWI lesion mask. The initial DWI lesion was located by thresholding Apparent Diffusion Coefficient (ADC) at 600∙10-6 mm2/sec. Due to the nature of thresholding, the ADC mask overestimates the DWI lesion volume and consequently we initialized level-set algorithm on DWI image with ADC mask as prior knowledge. Combining the PWI and inverted DWI mask then yield the PWI-DWI mismatch mask. Four expert raters manually outlined the penumbra for each patient. To quantify the geometric and volumetric similarity between APS and manually outlined penumbra we used volumetric correlation and Dice coefficient (DC). The performance of classification was evaluated in terms of sensitivity and specificity. Results: Penumbra volumes determined by APS demonstrated excellent correlation with the manually outlined penumbra volumes R2 = 0.93, with a mean difference of 2.1 mL. A geometrical comparison of penumbra masks resulted in a median DC of 0.70 and the median sensitivity was 73 % and the median specificity was 97 %. The median penumbra mask processing time per patient was 21.4 seconds. Conclusion: Combining the automatically generated PWI and DWI lesion mask yields a fast, robust estimate of the PWI-DWI mismatch. The APS algorithm demonstrates exceptional agreement with manually outlined mismatch masks. Requiring no user intervention, this algorithm allows fast, reliable identification of tissue-at-risk, and we speculate this feature may aid clinical decision-making, thus reducing the critical door-to-needle time period.

Originalsprog	Engelsk
Publikationsdato	1 feb. 2012
Status	Udgivet - 1 feb. 2012

Citationsformater

@conference{87992e92f7b147bbbc5c6732b7ba0f4e,

title = "A Fully Automated Penumbra Segmentation Tool",

abstract = "Introduction: Perfusion- and diffusion weighted MRI (PWI/DWI) is widely used to select patients who are likely to benefit from recanalization therapy. The visual identification of PWI-DWI-mismatch tissue depends strongly on the observer, prompting a need for software, which estimates potentially salavageable tissue, quickly and accurately. We present a fully Automated Penumbra Segmentation (APS) algorithm using PWI and DWI images. We compare automatically generated PWI-DWI mismatch mask to mask outlined manually by experts, in 168 patients. Method: The algorithm initially identifies PWI lesions by applying a connected component-labeling algorithm on thresholded Time-To-Peak maps (TTP > 4 seconds). Smooth PWI lesion borders are then identified by applying a level-set algorithm on the initial PWI lesion mask. The initial DWI lesion was located by thresholding Apparent Diffusion Coefficient (ADC) at 600∙10-6 mm2/sec. Due to the nature of thresholding, the ADC mask overestimates the DWI lesion volume and consequently we initialized level-set algorithm on DWI image with ADC mask as prior knowledge. Combining the PWI and inverted DWI mask then yield the PWI-DWI mismatch mask. Four expert raters manually outlined the penumbra for each patient. To quantify the geometric and volumetric similarity between APS and manually outlined penumbra we used volumetric correlation and Dice coefficient (DC). The performance of classification was evaluated in terms of sensitivity and specificity. Results: Penumbra volumes determined by APS demonstrated excellent correlation with the manually outlined penumbra volumes R2 = 0.93, with a mean difference of 2.1 mL. A geometrical comparison of penumbra masks resulted in a median DC of 0.70 and the median sensitivity was 73 % and the median specificity was 97 %. The median penumbra mask processing time per patient was 21.4 seconds. Conclusion: Combining the automatically generated PWI and DWI lesion mask yields a fast, robust estimate of the PWI-DWI mismatch. The APS algorithm demonstrates exceptional agreement with manually outlined mismatch masks. Requiring no user intervention, this algorithm allows fast, reliable identification of tissue-at-risk, and we speculate this feature may aid clinical decision-making, thus reducing the critical door-to-needle time period. ",

author = "Kartheeban Nagenthiraja and Ribe, {Lars Riisgaard} and Hougaard, {Kristina Dupont} and Leif {\O}stergaard and Kim Mouridsen",

year = "2012",

month = feb,

day = "1",

language = "English",

}

TY - ABST

T1 - A Fully Automated Penumbra Segmentation Tool

AU - Nagenthiraja, Kartheeban

AU - Ribe, Lars Riisgaard

AU - Hougaard, Kristina Dupont

AU - Østergaard, Leif

AU - Mouridsen, Kim

PY - 2012/2/1

Y1 - 2012/2/1

N2 - Introduction: Perfusion- and diffusion weighted MRI (PWI/DWI) is widely used to select patients who are likely to benefit from recanalization therapy. The visual identification of PWI-DWI-mismatch tissue depends strongly on the observer, prompting a need for software, which estimates potentially salavageable tissue, quickly and accurately. We present a fully Automated Penumbra Segmentation (APS) algorithm using PWI and DWI images. We compare automatically generated PWI-DWI mismatch mask to mask outlined manually by experts, in 168 patients. Method: The algorithm initially identifies PWI lesions by applying a connected component-labeling algorithm on thresholded Time-To-Peak maps (TTP > 4 seconds). Smooth PWI lesion borders are then identified by applying a level-set algorithm on the initial PWI lesion mask. The initial DWI lesion was located by thresholding Apparent Diffusion Coefficient (ADC) at 600∙10-6 mm2/sec. Due to the nature of thresholding, the ADC mask overestimates the DWI lesion volume and consequently we initialized level-set algorithm on DWI image with ADC mask as prior knowledge. Combining the PWI and inverted DWI mask then yield the PWI-DWI mismatch mask. Four expert raters manually outlined the penumbra for each patient. To quantify the geometric and volumetric similarity between APS and manually outlined penumbra we used volumetric correlation and Dice coefficient (DC). The performance of classification was evaluated in terms of sensitivity and specificity. Results: Penumbra volumes determined by APS demonstrated excellent correlation with the manually outlined penumbra volumes R2 = 0.93, with a mean difference of 2.1 mL. A geometrical comparison of penumbra masks resulted in a median DC of 0.70 and the median sensitivity was 73 % and the median specificity was 97 %. The median penumbra mask processing time per patient was 21.4 seconds. Conclusion: Combining the automatically generated PWI and DWI lesion mask yields a fast, robust estimate of the PWI-DWI mismatch. The APS algorithm demonstrates exceptional agreement with manually outlined mismatch masks. Requiring no user intervention, this algorithm allows fast, reliable identification of tissue-at-risk, and we speculate this feature may aid clinical decision-making, thus reducing the critical door-to-needle time period.

AB - Introduction: Perfusion- and diffusion weighted MRI (PWI/DWI) is widely used to select patients who are likely to benefit from recanalization therapy. The visual identification of PWI-DWI-mismatch tissue depends strongly on the observer, prompting a need for software, which estimates potentially salavageable tissue, quickly and accurately. We present a fully Automated Penumbra Segmentation (APS) algorithm using PWI and DWI images. We compare automatically generated PWI-DWI mismatch mask to mask outlined manually by experts, in 168 patients. Method: The algorithm initially identifies PWI lesions by applying a connected component-labeling algorithm on thresholded Time-To-Peak maps (TTP > 4 seconds). Smooth PWI lesion borders are then identified by applying a level-set algorithm on the initial PWI lesion mask. The initial DWI lesion was located by thresholding Apparent Diffusion Coefficient (ADC) at 600∙10-6 mm2/sec. Due to the nature of thresholding, the ADC mask overestimates the DWI lesion volume and consequently we initialized level-set algorithm on DWI image with ADC mask as prior knowledge. Combining the PWI and inverted DWI mask then yield the PWI-DWI mismatch mask. Four expert raters manually outlined the penumbra for each patient. To quantify the geometric and volumetric similarity between APS and manually outlined penumbra we used volumetric correlation and Dice coefficient (DC). The performance of classification was evaluated in terms of sensitivity and specificity. Results: Penumbra volumes determined by APS demonstrated excellent correlation with the manually outlined penumbra volumes R2 = 0.93, with a mean difference of 2.1 mL. A geometrical comparison of penumbra masks resulted in a median DC of 0.70 and the median sensitivity was 73 % and the median specificity was 97 %. The median penumbra mask processing time per patient was 21.4 seconds. Conclusion: Combining the automatically generated PWI and DWI lesion mask yields a fast, robust estimate of the PWI-DWI mismatch. The APS algorithm demonstrates exceptional agreement with manually outlined mismatch masks. Requiring no user intervention, this algorithm allows fast, reliable identification of tissue-at-risk, and we speculate this feature may aid clinical decision-making, thus reducing the critical door-to-needle time period.

M3 - Conference abstract for conference

ER -