Benchmarking CNN-based Models against Transformer-based Models for Abdominal Multi-Organ Segmentation on the RATIC Dataset

Bayer L, Bhat S, Maier A (2026)

Publication Language: English

Publication Type: Conference contribution, Conference Contribution

Publication year: 2026

City/Town: https://arxiv.org/abs/2603.18616

Conference Proceedings Title: BAIOSPHERE

Event location: Erlangen DE

URI: https://arxiv.org/abs/2603.18616

DOI: https://arxiv.org/abs/2603.18616

Open Access Link: https://arxiv.org/abs/2603.18616

Abstract

Accurate multi-organ segmentation in abdominal CT scans is important for computer-aided diag nosis, trauma assessment and treatment planning. While convolutional neural networks (CNNs) have long been the standard approach for medical image segmentation, recent transformer-based models claim superior performance. However, emerging evidence suggests that these advan tages may diminish when CNN baselines are carefully optimized. In this exploratory study, we benchmark three transformer-based models– UNETR, SwinUNETR, and UNETR++– against the state-of-the-art CNN-based model SegResNet on the recently introduced RATIC abdominal trauma dataset. We evaluate both segmentation accuracy and computational efficiency to provide a balanced comparison relevant for clinical deployment. Our results show that SegResNet out performs all three transformer-based models on this dataset, confirming that well-adjusted CNNs remain highly competitive. Nevertheless, UNETR++ achieves near-comparable accuracy while of fering substantially higher efficiency, indicating that certain transformer architectures may still be advantageous in specific scenarios. In contrast, SwinUNETR underperforms more clearly. These findings highlight the importance of rigorous and fair benchmarking when assessing new model families for clinical imaging tasks. Building on these benchmarking results, future research could explore the integration of these segmentation pipelines into clinical workflows, such as automated trauma grading and specific injury detection. To support future research, we provide our complete benchmarking pipeline as an open resource.

Authors with CRIS profile

Sheethal Bhat Lehrstuhl für Informatik 14 (Bild- und Sprachverarbeitung) (LME) Andreas Maier Lehrstuhl für Informatik 5 (Mustererkennung)

How to cite

APA:

Bayer, L., Bhat, S., & Maier, A. (2026). Benchmarking CNN-based Models against Transformer-based Models for Abdominal Multi-Organ Segmentation on the RATIC Dataset. In BAIOSPHERE. Erlangen, DE: https://arxiv.org/abs/2603.18616.

MLA:

Bayer, Lukas, Sheethal Bhat, and Andreas Maier. "Benchmarking CNN-based Models against Transformer-based Models for Abdominal Multi-Organ Segmentation on the RATIC Dataset." Proceedings of the BAIOSPHERE, Erlangen https://arxiv.org/abs/2603.18616, 2026.

BibTeX: Download