In Part 1 of this series, we discussed the basics of AI terminology – algorithms, machine learning, and deep learning – with examples from everyday life. Here in Part 2, we delve further into additional concepts of AI with an eye on how they pertain to radiology.
A neural network (also called an artificial neural network or ANN) is simply math and code. The neural network uses this set of mathematical algorithms to recognize patterns in a manner that loosely replicates the decision-making process of the human brain – basically clustering and labeling data.
A node in the neural network, similar to a neuron in the human brain, is a place where mathematical computation occurs in response to stimuli. In the neural network, that stimulus is inputted data. The node merges data with coefficients or weights to best solve the task instructed by the algorithm. The outcome either activates the node, passing data through to the next layer, or it adjusts and repeats the cycle.
Shallower single-hidden-layer neural networks are characterized by:
- Input layer – data in
- Hidden layer – computations and weighting made in nodes
- Output layer – adjusted data
A deep neural network1 or DNN is an artificial neural network with multiple layers of mathematical equations and millions of connections and parameters which get trained and strengthened based on the desired output.
A DNN contains a stacked hierarchy (rather than linear) of algorithms in a network having more than three layers. Thus, there is more than one hidden layer, and there could be many. Each progressive layer of nodes learns from the output of the previous layer. This multistep process is essential to pattern recognition. The deeper the layers go, the more aptly the DNN can aggregate information to cluster and label data.
A DNN is capable of discerning structure in raw, otherwise unstructured data. While the field of computer science has begun to build massive databases of labeled data, most data in the world around us is unlabeled. The job of the DNN is to discover similarities and anomalies in unorganized data. That is accomplished through training.
You want the truth?
The term “ground truth” originated in meteorology2, referring to independent confirmation of information obtained remotely. For example, a human spotting a tornado on-site, that had previously been detected on doppler radar.
Ground truth training data, in the context of radiology, refers to millions of images reconstructed by (ideally) filtered back projection to faithfully represent the scanned objects.
In supervised training3, a DNN runs inputted data through its layers of nodes. It continually compares output to established ground truths, updating its weights to formalize the structure of the new data until it can no longer reduce error.
We mentioned that one of the purposes of deep learning is clustering or the classification of data. Supervised learning is the transfer of human knowledge to the training dataset in the form of labels so that the neural network can learn to classify new inputs.
Human labeling identifies ground truths. The architecture of the algorithm identifies weights, which are essentially the guesswork applied to the input.
You might visualize the three key functions of the DNN this way:
- Scoring input
Input (new data) * weight = DNNs guess at the classification of data.
- Calculating loss
The difference between the ground truth and the DNN’s guess is an error (ground truth – guess = error). The DNN is saying, “I analyzed this data and made an assumption about what it is. Did I get it right?”
- Updating the model
If, compared to the ground truth, the answer is no, the DNN adjusts each weight to the extent it contributed to the errored assumption. Then the first two steps are repeated in a continuous feedback loop that rewards weights supporting correct guesses.
With deep learning, a neural network can detect similarities without labels. So, why bother to create a supervised learning structure? Because supervised learning has great potential to produce accurate models. It is the reverse of the old “garbage in/garbage out” adage. With supervised learning, a training dataset of high-quality ground truths generates highly accurate outputs.
Unsupervised learning4 is the ability of the machine to learn without labels, detecting similarities to determine the grouping (cluster or classification). Unsupervised learning can create accurate guesses, but masses of training data are necessary. In fact, the more data that is supplied during the DNN’s learning phase, the more accurate the models will be.
A DNN tackles unsupervised training differently than the supervised method outlined above. Each node layer in the DNN learns features of the inputted data by repeatedly trying to reconstruct it. Without ground truth, the DNN takes this alternate route to minimize disparities between inputted data and its own reconstruction or guess, to self-generate a label.
Working in harmony
A high-performance DNN is created by combining supervised and unsupervised training. The neural network is first trained on an extensive dataset of labeled ground truths. This model is then provided access to a larger mass of unstructured data. The additional input fine-tunes the model. In essence, a poor algorithm trained on a great deal of data will outperform a good algorithm trained on only a small amount of data. Bringing the good algorithm together with plenty of data enhances the DNN's learning capacity.
What does it mean to radiology now and into the future?
The intent of deep learning as applied to radiographic image reconstruction is improved prediction. Deep learning image reconstruction identifies similarities and anomalies in images. Both aspects are vital in detecting structural abnormalities in scans. Artificial intelligence is propelling radiology toward fewer surprises, which equates to more accurate diagnosis and more effective treatment planning. The goal of employing artificial intelligence in radiology is to arrive at the point of least error, as efficiently as possible.
In “Deep Learning: A Primer for Radiologists” published in RadioGraphics, Chartrand et al (2017) summed it up succinctly, “The introduction of deep learning techniques in radiology will likely assist radiologists in a variety of diagnostic tasks. Familiarity with the concepts, strengths, and limitations of computer-assisted techniques based on deep learning is critical to ensure optimal patient care.”5
- Deep learning (deep neural network). Tech Target. https://searchenterpriseai.techtarget.com/definition/deep-learning-deep-neural-network July 12, 2019.
- Ground Truth. Technopedia. https://www.techopedia.com/definition/32514/ground-truth July 12, 2019.
- Supervised Machine Learning: Classification. Towards Data Science. https://towardsdatascience.com/supervised-machine-learning-classification-5e685fe18a6d July 12, 2019.
- Supervised vs. Unsupervised Learning. Towards Data Science. https://towardsdatascience.com/supervised-vs-unsupervised-learning-14f68e32ea8d July 12, 2019.
- Deep Learning: A Primer for Radiologists. RadioGraphics. https://pubs.rsna.org/doi/full/10.1148/rg.2017170077 July 12, 2019.