Efficient Cardiac Image Segmentation with Compressed Vision Transformers and Post-training Quantization

Assia Boukhamla; Tamer Abderrahmane Lafia; Nabiha Azizi; Samir Brahim Belhaouari

doi:10.1007/978-981-96-6103-9_5

Efficient Cardiac Image Segmentation with Compressed Vision Transformers and Post-training Quantization

Assia Boukhamla^*, Tamer Abderrahmane Lafia, Nabiha Azizi, Samir Brahim Belhaouari

^*Corresponding author for this work

CSE Information & Computing Technology

Research output: Chapter in Book/Report/Conference proceeding › Chapter › peer-review

Abstract

The high prevalence of cardiovascular diseases (CVDs) worldwide requires accurate diagnostic imaging, particularly through magnetic resonance imaging (MRI). The framework includes preprocessing for region-of-interest segmentation via ViTs, followed by PTQ to reduce model size while maintaining segmentation accuracy. Using a small calibration dataset, we apply PTQ to compress the ViT, significantly reducing storage requirements and latency without compromising precision. Experimental results indicate that Float16 quantization achieves an optimal balance between compression rate and segmentation accuracy, demonstrating the feasibility of ViTs for real-time applications.

Original language	English
Title of host publication	Lecture Notes in Computational Vision and Biomechanics
Publisher	Springer Science and Business Media B.V.
Pages	55-69
Number of pages	15
DOIs	https://doi.org/10.1007/978-981-96-6103-9_5
Publication status	Published - 20 Jul 2025

Publication series

Name	Lecture Notes in Computational Vision and Biomechanics
Volume	40
ISSN (Print)	2212-9391
ISSN (Electronic)	2212-9413

Keywords

Cardiac image segmentation
Deep model compression
Post-training quantization
Vision transformer

Access to Document

10.1007/978-981-96-6103-9_5

Cite this

Boukhamla, A., Lafia, T. A., Azizi, N., & Belhaouari, S. B. (2025). Efficient Cardiac Image Segmentation with Compressed Vision Transformers and Post-training Quantization. In Lecture Notes in Computational Vision and Biomechanics (pp. 55-69). (Lecture Notes in Computational Vision and Biomechanics; Vol. 40). Springer Science and Business Media B.V.. https://doi.org/10.1007/978-981-96-6103-9_5

@inbook{92514f986ddf4bd28804d11573b988fd,

title = "Efficient Cardiac Image Segmentation with Compressed Vision Transformers and Post-training Quantization",

abstract = "The high prevalence of cardiovascular diseases (CVDs) worldwide requires accurate diagnostic imaging, particularly through magnetic resonance imaging (MRI). The framework includes preprocessing for region-of-interest segmentation via ViTs, followed by PTQ to reduce model size while maintaining segmentation accuracy. Using a small calibration dataset, we apply PTQ to compress the ViT, significantly reducing storage requirements and latency without compromising precision. Experimental results indicate that Float16 quantization achieves an optimal balance between compression rate and segmentation accuracy, demonstrating the feasibility of ViTs for real-time applications.",

keywords = "Cardiac image segmentation, Deep model compression, Post-training quantization, Vision transformer",

author = "Assia Boukhamla and Lafia, \{Tamer Abderrahmane\} and Nabiha Azizi and Belhaouari, \{Samir Brahim\}",

note = "Publisher Copyright: {\textcopyright} The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.",

year = "2025",

month = jul,

day = "20",

doi = "10.1007/978-981-96-6103-9\_5",

language = "English",

series = "Lecture Notes in Computational Vision and Biomechanics",

publisher = "Springer Science and Business Media B.V.",

pages = "55--69",

booktitle = "Lecture Notes in Computational Vision and Biomechanics",

address = "Germany",

}

Boukhamla, A, Lafia, TA, Azizi, N & Belhaouari, SB 2025, Efficient Cardiac Image Segmentation with Compressed Vision Transformers and Post-training Quantization. in Lecture Notes in Computational Vision and Biomechanics. Lecture Notes in Computational Vision and Biomechanics, vol. 40, Springer Science and Business Media B.V., pp. 55-69. https://doi.org/10.1007/978-981-96-6103-9_5

Efficient Cardiac Image Segmentation with Compressed Vision Transformers and Post-training Quantization. / Boukhamla, Assia; Lafia, Tamer Abderrahmane; Azizi, Nabiha et al.
Lecture Notes in Computational Vision and Biomechanics. Springer Science and Business Media B.V., 2025. p. 55-69 (Lecture Notes in Computational Vision and Biomechanics; Vol. 40).

Research output: Chapter in Book/Report/Conference proceeding › Chapter › peer-review

TY - CHAP

T1 - Efficient Cardiac Image Segmentation with Compressed Vision Transformers and Post-training Quantization

AU - Boukhamla, Assia

AU - Lafia, Tamer Abderrahmane

AU - Azizi, Nabiha

AU - Belhaouari, Samir Brahim

PY - 2025/7/20

Y1 - 2025/7/20

N2 - The high prevalence of cardiovascular diseases (CVDs) worldwide requires accurate diagnostic imaging, particularly through magnetic resonance imaging (MRI). The framework includes preprocessing for region-of-interest segmentation via ViTs, followed by PTQ to reduce model size while maintaining segmentation accuracy. Using a small calibration dataset, we apply PTQ to compress the ViT, significantly reducing storage requirements and latency without compromising precision. Experimental results indicate that Float16 quantization achieves an optimal balance between compression rate and segmentation accuracy, demonstrating the feasibility of ViTs for real-time applications.

AB - The high prevalence of cardiovascular diseases (CVDs) worldwide requires accurate diagnostic imaging, particularly through magnetic resonance imaging (MRI). The framework includes preprocessing for region-of-interest segmentation via ViTs, followed by PTQ to reduce model size while maintaining segmentation accuracy. Using a small calibration dataset, we apply PTQ to compress the ViT, significantly reducing storage requirements and latency without compromising precision. Experimental results indicate that Float16 quantization achieves an optimal balance between compression rate and segmentation accuracy, demonstrating the feasibility of ViTs for real-time applications.

KW - Cardiac image segmentation

KW - Deep model compression

KW - Post-training quantization

KW - Vision transformer

UR - https://www.scopus.com/pages/publications/105012025495

U2 - 10.1007/978-981-96-6103-9_5

DO - 10.1007/978-981-96-6103-9_5

M3 - Chapter

AN - SCOPUS:105012025495

T3 - Lecture Notes in Computational Vision and Biomechanics

SP - 55

EP - 69

BT - Lecture Notes in Computational Vision and Biomechanics

PB - Springer Science and Business Media B.V.

ER -

Efficient Cardiac Image Segmentation with Compressed Vision Transformers and Post-training Quantization

Abstract

Publication series

Keywords

Access to Document

Other files and links

Fingerprint

Cite this