top of page

AI microscopy glossary

Artificial intelligence – System of group of computer powered systems which mimic human intelligence in some way.

Machine learning – Group of algorithms and statistical models that can derive rules and patterns from data. The derived rules can then be used to process data in an automated way. Machine learning systems continuously adapt as new training data introduced and incorporated into the models.

Deep learning – Subset of machine learning algorithms.

AI Microscopy / smart microscopy – Microscopy done with systems that can mimic human intelligence. For example, such imaging systems would be able to auto focus, detect cells of interest, track each of them over long periods of time while (in real time) adjusting all the imaging parameters (hardware and software) to ensure images are of high quality.

Unsupervised (learning) – When the learning system is provided with unlabeled data only. This type of learning system can infer patterns / rules / groups by exploring all possible options. Typically, this type of approach requires larger amounts of training data and is more compute intensive versus supervised learning.

Supervised (learning) – The learning system is provided with labelled data in the form of Ground Truth. The learning system then learns the (typically) non-linear relationships between ground truth a raw data.

Ground truth / labelled / annotated data – Data that has been categorized by a human (typically) or by an automated system (could be an AI agent).

Raw data / unlabeled – Data that lacks categorization.

(Artificial) neural networks – a system composed of multiple algorithms which mimic the way the human brain works. Such systems are composed of neuron representations often organized in multiple connected layers. Individual artificial neurons can be influenced by upstream input signals, can process the signal, and pass the processed signal to downstream neurons, thus influencing them. Artificial neural networks, like real neural networks, dynamically adapt to the information flowing through them by weakening some neuron connections and strengthening others.

• AKA, ANNs.

Convolution – A mathematical operation in which a patch of input data (e.g. 3 x 3 pixels) are processed with a filter (e.g. gaussian) thus producing a new pixel with a resulting value. By processing a full image with said convolution one would create a feature map for the specific filter used (in this first example the gaussian filter feature map).

Convolutional (artificial) neural networks – This type of ANN uses convolutions to extract features from local regions of an input. Most CNNs contain a combination of convolutional, pooling and affine layers. CNNs are particularly popular in image related applications. The u-net is likely the most popular CNN architecture in use today in the biomedical field.

• AKA ConvNets, CNNs, CANNs.

bottom of page