Related Experiment Video
Updated: Jun 13, 2026
14:28
Substructure Analyzer: A User-Friendly Workflow for Rapid Exploration and Accurate Analysis of Cellular Bodies in Fluorescence Microscopy Images
Published on: July 15, 2020
TransNet-SAM2: A Transformer-Foundation Model Framework for Prompt-Free Segmentation of White Blood Cells in
Julius Bamwenda1, Mehmet Siraç Özerdem1, Orhan Ayyildiz2
1Electrical & Electronics Engineering Department, Engineering Faculty, Dicle University, 21280 Diyarbakır, Türkiye.
Diagnostics (Basel, Switzerland)
|June 12, 2026
Summary
This study introduces TransNet-SAM2, a novel deep learning framework for automated white blood cell (WBC) segmentation in blood smear images. The model achieves high accuracy, outperforming existing methods and offering a robust foundation for computational hematology.
Area of Science:
- Computational Hematology
- Medical Image Analysis
- Deep Learning
Background:
- Accurate white blood cell (WBC) segmentation is crucial for computational hematology tasks.
- Challenges include staining variability, complex cell morphology, and overlapping structures.
- Existing methods struggle with reliability and annotation requirements.
Purpose of the Study:
- To develop a robust, fully automated segmentation framework for WBCs in microscopic blood smear images.
- To provide a reliable foundation for downstream computational hematology analyses.
- To address limitations of current segmentation techniques.
Main Methods:
- Proposed TransNet-SAM2, a hybrid deep learning architecture integrating a Swin Transformer backbone with a foundation-model-based decoder.
- Employed hierarchical feature extraction and a feature adaptation module for multi-scale contextual representation.
- Utilized a weakly supervised self-training strategy for improved generalization with reduced annotation needs.
- Incorporated prompt-free segmentation by directly injecting fused features into the SAM2 mask decoder.
Main Results:
- TransNet-SAM2 achieved a Dice coefficient of 0.95 ± 0.01 and IoU of 0.90 on internal testing.
- Significantly outperformed U-Net, Mask R-CNN, and SAM2 (p < 0.05).
- Demonstrated robustness to domain shifts in cross-dataset evaluation (Dice: 0.91, IoU: 0.84).
- Ablation studies confirmed the contribution of each component, improving Dice by 6% over a CNN baseline.
Conclusions:
- The proposed TransNet-SAM2 framework offers a promising and scalable solution for WBC segmentation.
- The model shows high accuracy and robustness, even with overlapping cells and background clutter.
- Future work includes integration with classification and radiomics pipelines and validation on diverse clinical datasets.
