AssistDent, developed and patented by Manchester Imaging Limited, uses a range of machine
learning algorithms to identify regions of bitewing radiographs which are indicative of early enamelonly
interproximal caries. Detection of early enamel interproximal caries by the general dental
practitioner (GDP) is known to be difficult and yet once identified is at the stage when progress may
be prevented, or in some cases reversed, with measures such as cleaning or diet advice, fluoride
varnishes and resin infiltration.

A higher-grade caries that has penetrated beyond the enamel into
dentine can be detected with relative ease by the GDP, but preventative action is no longer possible.
Machine learning is a sub-discipline of Artificial Intelligence and is widely used in image analysis.
Rather than looking for pre-determined patterns in an image using a fixed algorithm, machine
learning captures knowledge from training examples of the features of interest and stores this in
mathematical models. These mathematical models are then used to detect regions which are
indicative of the trained features in previously unseen images.

For machine learning algorithms to offer clinical gains, the training examples need to be as close to a
recognised Gold Standard as possible. Any false positive examples in the training data will be
captured as true positives in the models and result in similar features being erroneously labelled
positive in the analysed images. Manchester Imaging has taken great care to ensure that data used
to train the AssistDent machine learning algorithms are of the highest possible standard. Five
internationally recognised maxillo-facial radiologists (experts in the analysis of radiographs of the
mouth region) were instructed to independently identify interproximal caries in a set of training
images. A consensus data set from these experts, deemed to be the Gold Standard, was then used to
train the AssistDent machine learning models. The result is a software tool which uses the combined
expertise from five authoritative maxillo-facial radiologists in its learning algorithm and is able to
detect the often very subtle patterns indicative of early enamel interproximal caries which can be
very difficult for GDPs to spot with accuracy.

The AssistDent model of machine learning uses “off-line” training, meaning the training exercise is
separate from the run-time execution of the software, and allows Manchester Imaging to train the
models using expert observers. Studies by Manchester Imaging and others have shown that GDPs
identify early caries with low sensitivity, finding approximately 30% to 40% compared to experts.
That is to say many early caries go undetected, because the signs on the bitewing radiograph are
subtle. Use of expert annotation means that AssistDent can detect early caries that may be missed
by GDPs. Manchester Imaging has embarked on an ambitious programme to acquire radiographs
from GDPs who use a range of x-ray capture equipment. These radiographs, together with analysis
by the Manchester Imaging panel of maxillo-facial experts, will ensure continual performance gains
as additional training data is incorporated.

An alternative model of machine learning is “on-line” learning. In this case human input into the
process at run-time continually adjusts and improves the models. This is often used in applications
where the intention is to achieve “human-like” performance such as face or object recognition. In
the case of caries detection, the intention is to improve on the general human performance, making
the performance by the GDP more like an expert maxillo-facial radiologist. On-line training by the
GDP would result in the models becoming increasingly trained to behave like the GDP and less like
the expert. An automated caries detection system is only worth having if it improves the GDP’s
performance.
The Manchester Imaging method of training constantly improves the models within AssistDent by
increasing the size of the expertly annotated training data set.