This project implements a U-Net model for biomedical image segmentation, designed to label neurons in electron microscopy images.
U-Net is a convolutional neural network (CNN) architecture that excels at pixel-level predictions, especially for segmentation tasks in medical imaging where spatial precision is critical.
Traditional CNNs lose spatial information during downsampling, which limits their usefulness for precise segmentation.
U-Net overcomes this by introducing skip connections between the encoder (contracting path) and decoder (expanding path), allowing the network to:
- Capture both context (what is in the image) and localization (where it is),
- Efficiently learn from a limited number of annotated biomedical images,
- Achieve state-of-the-art segmentation accuracy with relatively small datasets.
The dataset used comes from electron microscopy of neural tissue:
Arganda-Carreras et al. (2015). "Crowdsourcing the creation of image segmentation algorithms for connectomics." Frontiers in Neuroanatomy.
Link
The dataset consists of grayscale electron microscopy images and their corresponding segmentation masks. Each pixel is labeled as part of a neuron or background.
The model follows the U-Net design, comprising:
- Contracting Path (Encoder):
Multiple convolutional and max pooling layers that extract increasingly abstract image features while reducing spatial resolution. - Expanding Path (Decoder):
Upsampling and convolutional layers that reconstruct the segmentation mask, using skip connections to retain spatial detail from the encoder. - Final Output Layer:
A convolution layer that maps the feature maps to a binary segmentation mask.
- Loss Function: Binary cross-entropy, optimized for per-pixel classification.
- Optimizer: Adam optimizer for efficient gradient-based learning.
- Evaluation: Dice coefficient and visual inspection of predicted masks.
- Framework: Implemented using PyTorch for flexibility and GPU acceleration.
This notebook demonstrates how deep learning, specifically U-Net, can effectively perform neural tissue segmentation — a foundational step toward automated connectomics and neuroscience analysis.
It provides a practical implementation of supervised biomedical segmentation using a concise, interpretable architecture that has become a standard in the field.