U-Net: Convolutional Networks for Biomedical Image Segmentation (U-Net)

Implements a symmetric convolutional encoder-decoder architecture where the encoder progressively downsamples feature maps through repeated convolution and max-pooling operations, while the decoder upsamples through transposed convolutions. Skip connections concatenate encoder feature maps at each decoder level, preserving spatial detail lost during downsampling. This architecture enables pixel-level classification by combining coarse semantic information from deep layers with fine spatial information from shallow layers, allowing the network to learn both what and where to segment.

Applies learnable elastic deformations (random displacement fields) during training to artificially expand small biomedical datasets without requiring additional annotations. The method generates random displacement vectors on a coarse grid, interpolates them smoothly via B-splines, and applies the resulting deformation field to both input images and segmentation masks. This preserves anatomical realism (unlike naive rotation/scaling) by mimicking natural biological variation, enabling effective training on datasets with 30-100 annotated images by generating thousands of augmented variants per epoch.

Combines feature maps from multiple encoder depths during decoding by upsampling coarse feature maps via transposed convolutions and concatenating them with corresponding encoder skip connections. Each decoder block receives both upsampled features (containing semantic information from deeper layers) and skip-connected features (containing spatial detail from shallower layers), enabling the network to make segmentation decisions using both coarse context and fine detail. This multi-scale fusion is applied iteratively at 4-5 resolution levels, progressively refining segmentation predictions from coarse to fine.

Trains the entire encoder-decoder network end-to-end using pixel-level cross-entropy loss (or weighted variants) computed between predicted segmentation masks and ground-truth annotations. The loss is backpropagated through all layers simultaneously, enabling joint optimization of feature extraction (encoder) and spatial refinement (decoder). Supports weighted cross-entropy to handle class imbalance (e.g., background >> foreground in medical images), where each pixel's loss contribution is scaled by class frequency weights, allowing the network to learn meaningful segmentations despite skewed class distributions.

Enables inference on images larger than the training input size (e.g., 572×572 training → 1024×1024 inference) by decomposing large images into overlapping tiles, processing each tile independently through the network, and stitching predictions together. The fully convolutional architecture (no fully-connected layers) allows variable input sizes, and overlapping tiles reduce boundary artifacts. This approach extends the model to handle clinical images of arbitrary dimensions without retraining, though it introduces computational overhead and potential stitching artifacts at tile boundaries.

Implements standardized preprocessing for medical images including intensity normalization (zero-mean, unit-variance per image), histogram equalization for contrast enhancement, and optional Gaussian filtering for noise reduction. Preprocessing is applied consistently to both training and inference data, ensuring model robustness to imaging variations across different scanners, acquisition protocols, and patient populations. The pipeline is typically implemented as a preprocessing step before model input, enabling the network to focus on learning segmentation patterns rather than handling raw intensity variations.