Related Experiment Videos
Benchmarking Deep Segmentation Architectures for Histopathological Nuclei Segmentation: A Controlled Study of
1Department of Artificial Intelligence and Machine Learning, Kayseri University, Kayseri, Turkey. abdulkadirgulsen@kayseri.edu.tr.
Journal of Imaging Informatics in Medicine
|August 11, 2026
Summary
This study benchmarks deep learning models for nuclei segmentation in histopathology. Pretrained Transformer models excel in accuracy, but CNNs offer competitive performance with lower computational costs, emphasizing a balance of factors for model selection.
Area of Science:
- Computational pathology
- Deep learning for medical image analysis
- Histopathology image segmentation
Background:
- Accurate nuclei segmentation is crucial for quantitative histopathological analysis and clinical applications like cancer grading.
- Deep learning has advanced biomedical image segmentation, but direct comparisons of CNN, Transformer, and hybrid architectures are limited.
- A controlled benchmark is needed to evaluate these architectures under harmonized conditions for nuclei segmentation.
Purpose of the Study:
- To present a controlled and reproducible benchmark of eight representative deep learning architectures for histopathological nuclei segmentation.
- To compare the performance of Convolutional Neural Network (CNN), Transformer-based, and hybrid CNN-Transformer models.
- To assess the impact of ImageNet pretraining on segmentation performance across different architectures.
Main Methods:
- Evaluated eight architectures (four CNNs, two Transformers, two hybrid) on MoNuSeg and CoNSeP datasets.
- Ensured identical preprocessing, augmentation, optimization, and evaluation metrics for all models.
- Quantified performance using Dice, IoU, BF1, HD95 metrics, cross-validation, statistical testing, and computational efficiency measures.
Main Results:
- Pretrained Transformer models (SegFormer-B2, UPerNet) achieved the highest accuracy (Dice, IoU, BF1 scores).
- CNN models (U-Net, U-Net++) provided competitive performance at significantly lower computational cost.
- ImageNet pretraining yielded the most benefit for Transformer and hybrid architectures.
Conclusions:
- No single architectural paradigm is universally superior for nuclei segmentation.
- Model selection in computational pathology requires balancing accuracy (overlap and boundary), robustness, and computational efficiency.
- The choice of architecture should be guided by specific application requirements and resource constraints.