Paired microscopy images and binary masks resized to 256×256
Computer Vision · 19 MAR 2025
Biomedical Segmentation with U-Net
A pixel-level biomedical segmentation workflow that builds a U-Net from scratch, prepares microscopy images and masks, applies synchronized augmentation, and evaluates predictions with Dice and IoU.
TensorFlow / U-Net / Dice / IoU
PROJECT VISUAL / VERIFIEDTL;DR / KEY SIGNALS
Encoder-decoder network with skip connections
Thirty-epoch training with checkpointing and adaptive learning rate
Data preparation
Raw microscopy images and their binary masks are loaded as grayscale PNG files, resized to 256×256, and normalized to the 0–1 range.
A tf.data pipeline applies matching random flips and rotations to each image-mask pair, then shuffles, batches, and prefetches the samples efficiently.
U-Net architecture
The contracting path uses repeated convolution blocks and pooling to capture context. The bottleneck encodes the deepest representation, while transposed convolutions rebuild spatial resolution.
Skip connections concatenate encoder features with decoder features, preserving the localization detail required for precise biomedical boundaries. A sigmoid layer produces the final pixel-wise probability map.
Training and evaluation
The model is optimized with Adam and binary cross-entropy over thirty epochs. ModelCheckpoint preserves the strongest version, while ReduceLROnPlateau lowers the learning rate when validation progress stalls.
Loss, Dice, and Intersection-over-Union reveal convergence and generalization. Final visual comparisons place source images, reference masks, probability maps, and predicted masks side by side.
FULL DOCUMENTATION
Open detailed PDFGo deeper into the complete project.
Open the original associated PDF for the full methodology, code, analysis, and project evidence.