Authors - Nailfaaz, Wahyono Abstract - Accurate segmentation of anatomical structures in chest radiography (CXR) is critical for automated diagnosis. While CNNs achieve high regional overlap, they struggle with precise organ boundaries due to X-ray projection artifacts. This study systematically evaluates 32 encoder–decoder configurations combining U-Net and DeepLabV3+ with ResNet, MobileNet, and EfficientNet families to isolate Conditional Random Field-as-RNN (CRF-as-RNN) refinement impact on boundary quality. Results show U-Net outperforms DeepLabV3+ in preserving anatomical details. Crucially, a ”capacity threshold” is identified: CRF integration significantly reduces Hausdorff distances for lightweight models but yields diminishing returns for high-capacity backbones where baseline topology is already optimal.
