ML is a branch of AI that enables computers to learn from data by identifying patterns in data. By recognizing these patterns from previous examples, ML models can make predictions or decisions on new, unseen cases without relying on fixed rules for every situation.

A helpful analogy is to liken this process to how a dental student learns to identify interproximal caries on a bitewing radiograph. A student is not taught a million rigid rules describing every possible presentation of decay. Instead, they learn by reviewing hundreds of examples of both sound and carious teeth. Through this exposure, the student intuitively learns to recognize the subtle patterns, shapes, and changes in density that signify decay, and ultimately to diagnose whether a given region represents caries. An ML model learns in a conceptually similar way, discerning complex patterns from data that would be impossible for a human to program explicitly and using those patterns to classify new cases or make predictions. This ability to learn from experience makes ML models flexible and capable of handling the vast variation inherent in clinical data.

The traditional ML process can be understood as a pipeline with several key stages, as illustrated in Fig. 1 . The process begins with data collection , which is the first and most crucial step. This can be any form of information, from structured data in patient charts (age, probing depths, and medical history) to unstructured data like clinical notes or images.

A simplified flowchart of the traditional ML process.

The next stage is feature engineering , where meaningful variables, called features, are identified or derived from the raw data. In this step, domain experts play a key role in selecting features that are clinically relevant and informative for the task, or in transforming and creating new input variables that help the model learn more effectively from the data.

These features are then passed into the ML model training phase. In this stage, ML models learn to identify patterns within the data using algorithms such as linear regression, support vector machines (SVM), random forests, or clustering methods. The choice of algorithm and how the model learns depend on the type of information available. In some situations, the process is guided by examples with known diagnoses, allowing the model to learn how to classify new cases. In other cases, the model explores the data without known outcomes, uncovering natural groupings or trends.

ML can be divided into 2 types depending on whether each case includes a known diagnosis: supervised learning and unsupervised learning.

Supervised learning is the most common approach in medical diagnostics. In this paradigm, the algorithm learns from a dataset that has been labeled with the correct answers. The term supervised is used because an expert effectively supervises the algorithm by providing the ground truth for each data point. For example, to train an algorithm to detect periapical lesions, it would be fed hundreds of periapical radiographs, each one meticulously labeled by an endodontist as either lesion present or lesion absent . The algorithm’s task is to learn the mapping between the image and its corresponding label, so it can later classify new, unlabeled images.

Unsupervised learning , in contrast, works with data that has not been labeled. The algorithm is not provided with any correct answers; instead, its goal is to explore the data and find hidden structures or patterns on its own. A common application is clustering. For example, a researcher could feed an unsupervised algorithm the clinical records of hundreds of patients with temporomandibular disorders (TMD). The algorithm might automatically group these patients into 3 distinct clusters based on their symptoms, demographics, and treatment responses, potentially revealing previously unrecognized phenotypes of the disorder that could lead to more targeted therapies.

DL is a powerful and advanced subfield of ML that has been the driving force behind the most recent breakthroughs in AI, from self-driving cars to voice assistants. It uses complex architectures called deep neural networks , which are inspired by the intricate web of neurons and synapses in the human brain. These networks consist of many layers of interconnected nodes, with each layer building upon the output of the previous one to learn progressively more complex representations of the data.

An analogy can help clarify the difference. If a traditional ML model is like a simple light switch, designed to perform a specific, well-defined function, then a DL network is like the complex electrical wiring of an entire smart home. It has multiple layers of processing that allow it to learn and perform far more sophisticated and nuanced tasks, making it exceptionally well-suited for interpreting complex data like medical images.

The single most important advantage of DL over traditional ML is its ability to perform automatic feature extraction . As discussed, the traditional ML pipeline relies on human experts to manually identify and engineer the features the model will use. This process can be time-consuming, subjective, and may fail to capture the full complexity of the data.

DL models bypass this manual step entirely. The deep neural network learns the most important features directly from the raw data. Let us revisit the periapical lesion example. With a traditional ML approach, a clinician would have to program the model to look for specific features like the shape of the radiolucency, the integrity of the lamina dura, and the width of the periodontal ligament space. With DL, a clinician simply provides the model with hundreds of labeled radiographs. Through its multilayered training process, the network learns on its own that these very characteristics, and potentially others that humans do not consciously recognize, are the critical features for making an accurate diagnosis. This automation not only saves time but also unlocks the potential for the model to discover novel patterns that may be imperceptible to the human eye.