Performance evaluation of lightweight convolutional neural networks on retinal lesion segmentation

Marlin Siebert; Philipp Rostalski

doi:10.1117/12.2611796

Performance evaluation of lightweight convolutional neural networks on retinal lesion segmentation

Titel in Übersetzung: Leistungsanalyse von leichtgewichtigen neuronalen Faltungsnetzen zur Segmentierung von Netzhautläsionen

Marlin Siebert, Philipp Rostalski

Abstract

Zusätzlich zur Entwicklung von automatischen Erkennungssystemen für diabetische Retinopathie (DR) auf der Grundlage von Deep Learning werden derzeit Anstrengungen zur Integration dieser Systeme in mobile Erkennungsgeräte unternommen, die schlanke und leichtgewichtige Algorithmen erfordern. Um den klinischen Einsatz zu ermöglichen, ist es außerdem wichtig, die Transparenz der Deep-Learning-Systeme zu erhöhen, welche in der Regel Black-Box-Modelle sind und daher keinen Einblick in ihre zugrunde liegende Logik bei der Entscheidungsfindung geben. Durch die Bereitstellung präziser Segmentierungsmasken für Läsionen, die mit dem Schweregrad der DR in Zusammenhang stehen, kann eine gute Intuition über die Entscheidungsfindung des Systems vermittelt werden. Damit also ein Algorithmus zur mobilen, transparenten Erkennung von DR gleichzeitig krankheitsbedingte Läsionen segmentieren und auf Edge-Geräten laufen kann, werden leichtgewichtige, effiziente Modelle benötigt, die in der Lage sind, präzise Segmentierungsmasken zu erzeugen. Dies steht jedoch im Widerspruch zu der allgemein hohen Komplexität von vollständig faltungsbasierten Architekturen, die üblicherweise für die Bildsegmentierung verwendet werden. In dieser Arbeit evaluieren wir sowohl die Laufzeit als auch die Segmentierungsleistung mehrerer leichtgewichtiger, vollständig faltungsbasierter Netze für die Segmentierung von DR-bezogenen Läsionen und bewerten ihr Potenzial zur Erweiterung mobiler DR-Bewertungssysteme im Hinblick auf eine verbesserte Transparenz. Zu diesem Zweck wird das U2-Netz herunterskaliert, um die Rechenlast durch eine Verringerung der Merkmalstiefe und die Anwendung von separierbaren Faltungen zu reduzieren, und die Verwendung von Deep-Ensembling sowie die Single- und Multi-Task-Inferenz evaluiert, um die Leistung zu verbessern sowie die Speicheranforderungen weiter zu reduzieren. Die experimentellen Ergebnisse mit dem U2-Net-S†-Ensemble zeigen eine gute Segmentierungsleistung bei geringem Speicherbedarf und angemessener Inferenzgeschwindigkeit und stellen somit einen vielversprechenden ersten Schritt in Richtung eines ganzheitlichen mobilen Diagnosesystems dar, welches sowohl eine präzise Läsionssegmentierung als auch eine DR-Schweregradklassifizierung bieten kann.

Titel in Übersetzung	Leistungsanalyse von leichtgewichtigen neuronalen Faltungsnetzen zur Segmentierung von Netzhautläsionen
Originalsprache	Englisch
Titel	Medical Imaging 2022: Computer-Aided Diagnosis
Redakteure/-innen	Khan M. Iftekharuddin, Karen Drukker, Maciej A. Mazurowski, Hongbing Lu, Chisako Muramatsu, Ravi K. Samala
Herausgeber (Verlag)	SPIE
Erscheinungsdatum	2022
DOIs	https://doi.org/10.1117/12.2611796
Publikationsstatus	Veröffentlicht - 2022

Strategische Forschungsbereiche und Zentren

Zentren: Center for Open Innovation in Connected Health (COPICOH)
Querschnittsbereich: Intelligente Systeme
Zentren: Zentrum für Künstliche Intelligenz Lübeck (ZKIL)

DFG-Fachsystematik

4.43-05 Bild- und Sprachverarbeitung, Computergraphik und Visualisierung, Human Computer Interaction, Ubiquitous und Wearable Computing

UN SDGs

Dieser Output leistet einen Beitrag zu folgendem(n) Ziel(en) für nachhaltige Entwicklung

Dokumente

10.1117/12.2611796

Zitieren

@inproceedings{93405fd84ef64f49a4d28cfa37f014b5,

title = "Performance evaluation of lightweight convolutional neural networks on retinal lesion segmentation",

abstract = "In addition to the recent development of deep learning-based, automatic detection systems for diabetic retinopathy (DR), efforts are being made to integrate those systems into mobile detection devices running on the edge requiring lightweight algorithms. Moreover, to enable clinical deployment it is important to enhance the transparency of the deep learning systems usually being black-box models and hence giving no insights into its reasoning. By providing precise segmentation masks for lesions being related to the severity of DR, a good intuition about the decision making of the diagnosing system can be given. Hence, to enable transparent mobile DR detection devices simultaneously segmenting disease-related lesions and running on the edge, lightweight models capable to produce fine-grained segmentation masks are required contradicting the generally high complexity of fully convolutional architectures used for image segmentation. In this paper, we evaluate both the runtime and segmentation performance of several lightweight fully convolutional networks for DR related lesion segmentation and assess its potential to extend mobile DR-grading systems for improved transparency. To this end, the U2-Net is downscaled to reduce the computational load by reducing feature size and applying depthwise separable convolutions and evaluated using deep model ensembling as well as single- and multi-task inference to improve performance and further reduce memory cost. Experimental results using the U2-Net-S† ensemble show good segmentation performance while maintaining a small memory footprint as well as reasonable inference speed and thus indicate a promising first step towards a holistic mobile diagnostic system providing both precise lesion segmentation and DR-grading.",

author = "Marlin Siebert and Philipp Rostalski",

year = "2022",

doi = "10.1117/12.2611796",

language = "English",

editor = "Iftekharuddin, {Khan M.} and Karen Drukker and Mazurowski, {Maciej A.} and Hongbing Lu and Chisako Muramatsu and Samala, {Ravi K.}",

booktitle = "Medical Imaging 2022: Computer-Aided Diagnosis",

publisher = "SPIE",

}

Performance evaluation of lightweight convolutional neural networks on retinal lesion segmentation. / Siebert, Marlin ; Rostalski, Philipp.
Medical Imaging 2022: Computer-Aided Diagnosis. Hrsg. / Khan M. Iftekharuddin; Karen Drukker; Maciej A. Mazurowski; Hongbing Lu; Chisako Muramatsu; Ravi K. Samala. SPIE, 2022.

Publikation: Kapitel in Büchern/Berichten/Konferenzbänden › Konferenzbeitrag › Begutachtung

TY - GEN

T1 - Performance evaluation of lightweight convolutional neural networks on retinal lesion segmentation

AU - Siebert, Marlin

AU - Rostalski, Philipp

PY - 2022

Y1 - 2022

N2 - In addition to the recent development of deep learning-based, automatic detection systems for diabetic retinopathy (DR), efforts are being made to integrate those systems into mobile detection devices running on the edge requiring lightweight algorithms. Moreover, to enable clinical deployment it is important to enhance the transparency of the deep learning systems usually being black-box models and hence giving no insights into its reasoning. By providing precise segmentation masks for lesions being related to the severity of DR, a good intuition about the decision making of the diagnosing system can be given. Hence, to enable transparent mobile DR detection devices simultaneously segmenting disease-related lesions and running on the edge, lightweight models capable to produce fine-grained segmentation masks are required contradicting the generally high complexity of fully convolutional architectures used for image segmentation. In this paper, we evaluate both the runtime and segmentation performance of several lightweight fully convolutional networks for DR related lesion segmentation and assess its potential to extend mobile DR-grading systems for improved transparency. To this end, the U2-Net is downscaled to reduce the computational load by reducing feature size and applying depthwise separable convolutions and evaluated using deep model ensembling as well as single- and multi-task inference to improve performance and further reduce memory cost. Experimental results using the U2-Net-S† ensemble show good segmentation performance while maintaining a small memory footprint as well as reasonable inference speed and thus indicate a promising first step towards a holistic mobile diagnostic system providing both precise lesion segmentation and DR-grading.

AB - In addition to the recent development of deep learning-based, automatic detection systems for diabetic retinopathy (DR), efforts are being made to integrate those systems into mobile detection devices running on the edge requiring lightweight algorithms. Moreover, to enable clinical deployment it is important to enhance the transparency of the deep learning systems usually being black-box models and hence giving no insights into its reasoning. By providing precise segmentation masks for lesions being related to the severity of DR, a good intuition about the decision making of the diagnosing system can be given. Hence, to enable transparent mobile DR detection devices simultaneously segmenting disease-related lesions and running on the edge, lightweight models capable to produce fine-grained segmentation masks are required contradicting the generally high complexity of fully convolutional architectures used for image segmentation. In this paper, we evaluate both the runtime and segmentation performance of several lightweight fully convolutional networks for DR related lesion segmentation and assess its potential to extend mobile DR-grading systems for improved transparency. To this end, the U2-Net is downscaled to reduce the computational load by reducing feature size and applying depthwise separable convolutions and evaluated using deep model ensembling as well as single- and multi-task inference to improve performance and further reduce memory cost. Experimental results using the U2-Net-S† ensemble show good segmentation performance while maintaining a small memory footprint as well as reasonable inference speed and thus indicate a promising first step towards a holistic mobile diagnostic system providing both precise lesion segmentation and DR-grading.

U2 - 10.1117/12.2611796

DO - 10.1117/12.2611796

M3 - Conference contribution

BT - Medical Imaging 2022: Computer-Aided Diagnosis

A2 - Iftekharuddin, Khan M.

A2 - Drukker, Karen

A2 - Mazurowski, Maciej A.

A2 - Lu, Hongbing

A2 - Muramatsu, Chisako

A2 - Samala, Ravi K.

PB - SPIE

ER -

Performance evaluation of lightweight convolutional neural networks on retinal lesion segmentation

Abstract

Strategische Forschungsbereiche und Zentren

DFG-Fachsystematik

UN SDGs

Dokumente

Fingerprint

Zitieren