Challenging Complexity Bias: Simpler Networks outperform Large Models in Multi-Organ Segmentation of Ultrasound Images

Sursoon, Ting Wu, Breezewen, Ziwei Nie

Keywords: Ultrasound segmentation, Model complexity, Pretraining strategies, Multiorgan segmentation, Computational efficiency

TL;DR: We found that simple CNNs outperform large deep learning models in Multi-Organ Segmentation of Ultrasound Images.

Abstract: The prevailing view maintains that large-parameter models excel at image feature extraction compared to small-parameter counterparts. However, our work challenges this ”complexity advantage” bias. This paper explores the utility of pretraining strategies for multi-organ segmentation in ultrasound images. Surprisingly, experimental results show that pretrained simple network architectures not only achieve higher segmentation accuracy than similarly pretrained complex networks but also offer significant advantages in computational efficiency and parameter scale. This insight provides new perspectives and solid evidence for developing efficient and lightweight ultrasound analysis tools suitable for clinical deployment.

Primary Subject Area: Segmentation

Secondary Subject Area: Foundation Models

Registration Requirement: Yes

Reproducibility: https://github.com/ZiweiNiexiaoer/MultiOrganUltrasoundImageSegmentation

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 266