Semantic Segmentation Masks of Cell Nuclei in Oral Leukoplakia Histopathological Images

This dataset provides binary semantic segmentation masks of cell nuclei for histopathological images of oral leukoplakia. The masks were generated by a suite of deep-learning segmentation models and ensemble strategies. What the masks represent. Each file is a single-channel (grayscale) PNG of 256×256 pixels. Foreground pixels (value 255) mark regions predicted as cell nuclei; background pixels (value 0) mark non-nuclei tissue. Predictions were obtained by applying a sigmoid activation to each model's output followed by thresholding at 0.5. Every mask preserves the filename of its corresponding source image, allowing direct one-to-one pairing with the original histopathological patches. How the masks were produced. The masks stem from ten heterogeneous segmentation architectures spanning three paradigms — fully-convolutional, hybrid CNN–Transformer, and attention-based networks: U-Net (ResNet-34) — baseline U-Net with scSE attention (ResNet-34) U-Net++ (EfficientNet-B6) U-Net++ with scSE (ResNeSt-50) MA-Net (MiT-B2) U-Net (ConvNeXt-Base) FPN (ResNet-50) TransUNet (ResNet-50 + Transformer encoder) SegFormer (MiT-B3) MaxViT with Deformable Large-Kernel Attention The individual predictions are further combined through five ensemble fusion rules: soft (mean) voting, hard (majority) voting, Dice-weighted soft voting, union (pixel-wise maximum), and intersection (pixel-wise minimum).