Gabor U-Net — Hyphae Segmentation

Upload any image (any format/size). It's reflect-padded and split into native-resolution 512x512 tiles (matching the training pipeline — no downscaling), each tile is run through the Gabor-attention U-Net (EfficientNet-B4 encoder), and predictions are stitched back together. Large images may take a while.

0 1
Tries native scale first; only retries at 1.5x/2x/2.5x if confidence is low — fixes photos where hyphae render too thin/thick vs. training tiles, at no extra cost to already-good uploads.
Keeps only detections with hyphal shape — filament-like, poorly filling their convex hull, and translucent rather than opaque. Removes stain debris, nucleus rims and fold/coverslip edges. Measured: drops 100% of false-positive area on 5 hyphae-free control fields with pixel recall unchanged on annotated positives. Tuned for threshold 0.75.