SENLA-ResUNet: Hybrid Residual U-Net with Attention for Nuclei Segmentation

SENLA-ResUNet is a novel deep learning architecture designed for accurate nuclei segmentation in biomedical images.
It enhances the standard Residual U-Net by incorporating multi-scale feature extraction, channel-wise recalibration, and global attention mechanisms, achieving state-of-the-art performance on challenging datasets like DSB2018 and TNBC.

📄 Reference: Research Paper

---

✨ Key Features

• 🔹 Residual U-Net Backbone – deep encoder-decoder with residual blocks for better gradient flow.
• 🔹 Encoder Attention Module – combines multi-scale Inception-SE blocks with Non-Local Attention.
• 🔹 Squeeze-and-Excitation (SE) blocks – adaptively recalibrate feature channels.
• 🔹 Attention Gates in Skip Connections – filter irrelevant features and focus on salient regions.
• 🔹 Global Context Awareness – captures long-range dependencies for robust boundary delineation.

---

🏗️ Architecture Overview

🔹 Overall SENLA-ResUNet

🔹 Encoder Attention Module

---

📊 Results

DSB-2018 Dataset

Method	IoU	Dice	Precision	Recall	Accuracy
U-Net	0.7277	0.7793	0.8775	0.8057	0.9650
ResUNet	0.7871	0.8014	0.8631	0.8991	0.9710
Attention U-Net	0.7688	0.7972	0.8447	0.8967	0.9680
Cellpose	0.6972	0.8166	0.8736	0.7747	0.8623
SENLA-ResUNet	0.7936	0.8172	0.8791	0.9084	0.9726

TNBC Dataset

Method	IoU	Dice	Precision	Recall	Accuracy
U-Net	0.3876	0.4735	0.8386	0.4188	0.9085
ResUNet	0.5145	0.6745	0.6657	0.6868	0.9096
Attention U-Net	0.4358	0.4577	0.4670	0.8706	0.8511
Cellpose	0.7380	0.8117	0.8283	0.8696	0.8564
SENLA-ResUNet	0.8332	0.8986	0.9353	0.8822	0.9582

✅ SENLA-ResUNet sets a new benchmark with superior IoU, Dice, Precision, Recall, and Accuracy.

---

📂 Dataset

• Data Science Bowl 2018 (DSB2018)
• TNBC Nuclei Segmentation Dataset

All images are resized to 128×128 pixels for training.

---

⚙️ Implementation Details

• Framework: TensorFlow + Keras
• Optimizer: Adam (lr=1e-4)
• Loss: Binary Cross-Entropy
• Batch Size: 16
• Early stopping + checkpointing used.
• Epochs: 20 max

---

🚀 Getting Started

Installation

git clone https://github.com/OVER-CODER/Nuclei-Segmentation.git
cd Nuclei-Segmentation

---

📌 Citation

If you use SENLA-ResUNet in your research, please cite:

@article{pandit2025senla,
  title={SENLA-ResUNet: A Hybrid Residual U-Net with Squeeze-Excitation and Non-Local Attention for Accurate Nuclei Segmentation},
  author={Pandit, Aryan and Mishra, Shivam and Vishwakarma, Amit and Kumar, Anil},
  year={2025},
  journal={arXiv preprint}
}