DIAMOND is a novel deep learning framework for medical image segmentation that integrates a dual-encoder hybrid backbone and a quasi-multimodal training paradigm. It achieves robust generalization, computational efficiency, and strong adaptability across heterogeneous datasets.
- Dual-Encoder Hybrid Backbone: Combines a convolutional encoder (for local feature extraction) with a Swin-Transformer encoder (for global context modeling).
- Residual Recursive Gated Convolution (RrgConv): A lightweight, portable module that stabilizes high-order spatial interactions with low computational overhead.
- Nested Attention System (NAS): Integrates spatial and channel attention within self-attention to enhance feature representation and focus.
- Quasi-Multimodal Training (QMM): Facilitates training across multiple datasets focused on the same lesion category, reducing annotation heterogeneity and improving cross-dataset generalization.
- Comprehensive Evaluation: Benchmarked on 10 public datasets and 33 state-of-the-art models, showing competitive or superior performance with reduced resource demands.
├── LICENSE
├── README.md
├── Diamond/
│ ├── Diamond_ECDC.py # Diamond Backbone Network CNN encoder – CNN decoder
│ ├── Diamond_ECDT.py # Diamond Backbone Network CNN encoder – Transformer decoder
│ ├── Diamond_ETDC.py # Diamond Backbone Network Transformer encoder – CNN decoder
│ ├── Diamond_ETDT.py # Diamond Backbone Network Transformer encoder – Transformer decoder
│ ├── Diamond_ECDC_NAS.py # Diamond_ECDC with MHSA replaced by Nested Attention System
│ ├── Diamond_ECDC_Rrg.py # Diamond_ECDC with Double Conv replaced by RrgConv in Bottleneck
│ ├── ...
│ ├── DoubleConv.py # Double Convolution Module
│ ├── DWSCov.py # Depthwise Separable Convolution
│ ├── HorBlock.py # The HorBlock of the HorNet
│ ├── MHSA.py # Multi Head Self Attention
│ ├── NAS.py # Nested Attention System
│ ├── RrgConv.py # Residual Recursive Gated Convolution
│ ├── Transition.py # 1x1 Convolution Transition Module
│ └── ...
├── Portability/
│ ├── DoubleConv.py # Double Convolution Module
│ ├── DWSCov.py # Depthwise Separable Convolution
│ ├── HorBlock.py # The HorBlock of the HorNet
│ ├── NAS.py # Nested Attention System
│ ├── RrgConv.py # Residual Recursive Gated Convolution
│ ├── U_Net_NAS.py # U-Net with Nested Attention System inserted in Bottleneck
│ └── U-Net_Rrg.py # U-Net with Double Conv replaced by RrgConv in Bottleneck
- DIAMOND achieves superior Dice and IoU metrics on multiple datasets.
- The QMM strategy consistently improves cross-dataset generalization.
- Lower computational and memory requirements compared to many existing models.
Sample qualitative results are provided in the results/sample_outputs/ directory.
This project is licensed under the MIT License.