Generic selectors
Exact matches only
Search in title
Search in content
Post Type Selectors

Convolutional neural networks explained

A convolutional neural network (CNN) is a neural net built for grids: a small kernel slides across an image, detects local patterns, and stacks those detections into a label, a box, or a mask. It is not a vision transformer. It is not “machine learning for medical imaging” as a field.

If you meant CNN vs ViT on medical imagesCNNs or ViT for medical imaging. If you meant what imaging ML returns (mark / mask / score)machine learning for medical imaging. If you meant how you check a trained modelmachine learning model evaluation.

PYCAD builds custom web DICOM viewers and imaging models. This page is the architecture, not a product.

The three stations

Layer What it does What you have after
Convolution A small kernel (often 3×3) multiplies a neighborhood, then steps. Each kernel is one pattern: an edge, a texture, later a shape A feature map: where that pattern fired
Pooling Usually max-pool: keep the strongest value in a 2×2 window. Shrinks the map Fewer pixels, some shift invariance. The exact location is fuzzier
Fully connected / head Takes the stacked maps and writes a vector: class scores, box coords, or (in a U-Net) a full-resolution mask The thing you evaluate

A dense net flattens the image into one long vector and throws the grid away. A CNN keeps the grid. Neighboring pixels stay neighbors. That is why a CNN is cheaper on a 512×512 film than a fully connected net of the same depth, and why it is the default backbone for 2D X-ray and 3D CT / MRI.

A non-linear activation (ReLU is the usual one) sits after the convolution so stacked layers can represent more than a single linear filter. Without it the whole stack collapses to one matrix multiply.

How it learns

Three steps, repeated on labeled images:

  1. Forward. Pixels go through the stack. The head guesses.
  2. Loss. A number for how wrong the guess was (cross-entropy for a class, Dice / cross-entropy for a mask).
  3. Backprop. That error is sent backward. Each kernel weight is nudged a little. An optimizer (SGD, Adam) owns the step size.

A new net guesses at random. After enough labeled studies the kernels become edge detectors, then texture detectors, then task detectors. How you trust that number on a new hospital’s scanner is model evaluation, not this page.

Landmark architectures

Name Year What it added Usual citation task
LeNet-5 1998 Conv + pool as a working stack Handwritten digits (MNIST)
AlexNet 2012 Depth, ReLU, GPU training. ImageNet top-5 error 15.3% vs 26.2% for the previous winner (Krizhevsky et al.) ImageNet classification
VGG 2014 Stacked 3×3 convs, simple and deep Recognition; a common feature backbone
GoogLeNet 2014 Inception: several kernel sizes in one block ImageNet, fewer parameters than a plain stack
ResNet 2015 Skip connections so a 50–150 layer net can train. ImageNet top-5 around 3.6% (He et al.) The default residual backbone

U-Net (2015, Ronneberger et al.) is the medical fork: an encoder–decoder with skips so the output is a mask the same size as the input, not a single class. That is the architecture most organ-segmentation papers still start from. The methods page is medical image segmentation.

On medical images

2D CNNs sit on X-ray, fundus, mammography, a single CT slice. 3D CNNs (and 3D U-Nets) sit on CT / MRI volumes because the finding has a z-extent. A 2D net on one slice will miss a nodule that is obvious two slices later.

A CNN does not sign a report. It returns a mark, a mask, or a score. The clinician uses or ignores it. That sentence is the job description on machine learning for medical imaging.

What this page is not

If a CNN has to run inside a viewer a clinic already uses, that is the imaging piece. Case studies.

We build custom medical imaging platforms — advanced DICOM viewers, AI segmentation, and the clinical systems around them.

Get in Touch

Copyright © 2026 PYCAD. All Rights Reserved.