ADCFormer: A hybrid model of Adaptive Depth-wise Convolution and Transformer for retinal macular edema segmentation in OCT images
Optical coherence tomography (OCT) is a non-invasive imaging technique. OCT technology has been widely used in the clinical setting for noninvasive diagnosis, etiological analysis, and therapeutic guidance of diabetic macular edema (DME), which has greatly improved the diagnosis and treatment of the...
Gespeichert in:
| Veröffentlicht in: | Biomedical signal processing and control Jg. 108; S. 107949 |
|---|---|
| Hauptverfasser: | , |
| Format: | Journal Article |
| Sprache: | Englisch |
| Veröffentlicht: |
Elsevier Ltd
01.10.2025
|
| Schlagworte: | |
| ISSN: | 1746-8094 |
| Online-Zugang: | Volltext |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | Optical coherence tomography (OCT) is a non-invasive imaging technique. OCT technology has been widely used in the clinical setting for noninvasive diagnosis, etiological analysis, and therapeutic guidance of diabetic macular edema (DME), which has greatly improved the diagnosis and treatment of the disease. However, the relative lack of contrast and the significant noise problem of OCT images pose great difficulties for ophthalmologists in recognizing the complete DME lesion area, thus severely limiting the accuracy of DME detection. In this research, a hybrid model of Adaptive Depth-wise Convolution and Transformer (ADCFormer) was proposed for accurate segmentation of DME lesion areas in OCT images. Our proposed ADCFormer model fully combines the advantages of convolutional neural network (CNN) and transformer. The model employs an Efficient Self-Attention mechanism (ESA) for global information modeling and Adaptive Depth-wise Convolution (ADWConv) to dynamically capture local effective information. Next, we design the Global Local Feature Fusion (GLFF) module to improve the modeling capability of multi-scale remote dependencies and local spatial information. In addition, we design a new feed-forward network, Residual Multi-scale Feed-forward Network (ReMS-FFN), in the encoder–decoder structural transformer block, which can fuse multi-scale feature information at different locations and levels to enhance the representation of global and local features. Finally, we tested ADCFormer on the Intraretinal Cystoid Fluid and the DUKE dataset, which outperformed nine state-of-the-art segmentation models with competitive results. The robustness and applicability of the model was further validated using Optima and hospital-collected datasets.
•A hybrid model ADCFormer can address the limitations of CNN and Transformer in OCT segmentation.•ADWConv, ESA, and GLFF modules can capture global and local dependencies effectively.•ReMS-FFN and CCFC enhance the fusion of global and local features in segmentation. |
|---|---|
| ISSN: | 1746-8094 |
| DOI: | 10.1016/j.bspc.2025.107949 |