This model for lung segmentation in chest X-ray images is based on a custom U-Net architecture enhanced with:
- ASPP (Atrous Spatial Pyramid Pooling) in the bottleneck to capture multi-scale context and anatomical structures of varying size
- SE (Squeeze-and-Excitation) blocks to enhance channel-wise attention and suppress irrelevant features such as ribs or background noise
- Dilated convolutions in the decoder to increase the receptive field without sacrificing spatial resolution
The model was trained on the COVID-19 Radiography Database and evaluated on a dedicated internal validation set.
It achieves a Dice score of 98.7%, demonstrating good performance in segmenting lung fields.