Applications of machine learning and deep learning in musculoskeletal medicine: a narrative review

Abstract

Artificial intelligence (AI), with its technologies such as machine perception, robotics, natural language processing, expert systems, and machine learning (ML) with its subset deep learning, have transformed patient care and administration in all fields of modern medicine. For many clinicians, however, the nature, scope, and resulting possibilities of ML and deep learning might not yet be fully clear. This narrative review provides an overview of the application of ML and deep learning in musculoskeletal medicine. It first introduces the concept of AI and machine learning and its associated fields. Different machine concepts such as supervised, unsupervised and reinforcement learning will then be presented with current applications and clinical perspective. Finally deep learning applications will be discussed. With significant improvements over the last decade, ML and its subset deep learning today offer potent tools for numerous applications to implement in clinical practice. While initial setup costs are high, these investments can reduce workload and cost globally. At the same time, many challenges remain, such as standardisation in data labelling and often insufficient validity of the obtained results. In addition, legal aspects still will have to be clarified. Until good analyses and predictions are obtained by an ML tool, patience in training and suitable data sets are required. Awareness of the strengths of ML and the limitations that lie within it will help put this technique to good use.

Keywords: Artificial intelligence; Machine learning; Orthopaedics; Reinforcement learning; Supervised learning; Traumatology; Unsupervised learning.

Fig. 1

Number of publications using machine learning in orthopaedics or traumatology has increased exponentially in the last 10 years. A PubMed search was conducted using the search terms “orthopaedics,” “traumatology,” and “machine learning.” The time range includes publications until 2022. While these techniques were hardly in use until 2015, the number of new publications has risen exponentially in the last decade and continues to do so

Fig. 2

Within the concept of artificial intelligence, deep learning forms the"intelligent"subset of machine learning

Fig. 3

Machine learning model to detect osteolysis in a plain knee radiograph. Labelled input radiographs of healthy and pathological knees are given to the system. The training model then decomposes these images into grey value pixels. The model defines edges at areas of transition from higher to lower grey values. These edges are then aligned with the already-learned anatomy of a healthy knee radiograph. This feature extraction process involves identifying and capturing essential healthy and pathological knee characteristics. Aberrant lines are finally marked and labelled as pathologic. For the model creation, this process is repeatedly iterated to improve the diagnostic value of the model further

Fig. 4

Three most common machine learning (ML) techniques. A machine learning model can be thought of as a complex web of interconnected nodes. Setting up an ML model involves two different kinds of data types: in the first step, training data are used to train the model. Once the model is set up in terms of its internal parameters, an unknown test dataset is used in a second step to validate the model. Finally, the model is used on new data. A Supervised learning problems can be sub-grouped into classification and regression techniques. In supervised learning, labelled data are used to train the model. This means that labelled input data are associated with a known outcome. The model is then trained on these data by an iterative process until fine-tuning of the model has been achieved. The model thus learns which features define the input data and how to identify them. This is done by applying weights, which represent numerical values assigned to connection nodes of the model. Weights determine the strength of these individual connections in the web of interconnected nodes and as such how strongly the output of a node influences another node’s input. Predictions made by supervised models can either be discrete or continuous. A model that produces discrete output data is a classification model (e.g., the result: tumour malignant or benign), and one that produces continuous output data is a regression model (e.g., the tolerable dose of a certain medication). B Unsupervised learning is used, e.g., clustering. Here, raw unlabelled data objects (on the left side) are provided as input. Training the model is also an iterative process. The results of unsupervised learning are often different clusters (as shown here with the non-overlapping geometrical shapes on the right). Clustering algorithms are used to assort the given data into groups that share common structures or patterns. C Reinforcement learning differs from supervised and unsupervised learning. In reinforcement learning, the model learns by the interactions between a decision maker/agent and its surrounding environment. The decision maker/agent selects an action according to its policy. Depending on the nature of the change in the environment, this action can be positive ("reward") which would reinforce the previous behaviour of the model, or negative ("punishment"). The goal of the model is to maximise its rewards

Fig. 5

Exemplary neural-network architecture to detect sarcoma in a conventional knee radiograph. Adapted from Schulz et Behnke [89] with permission. In this example, four layers are superimposed: 1. The computer identifies pixels of light and dark. 2. The computer learns to identify edges and simple shapes. 3. The computer learns to identify more complex shapes and objects and integrates them into the notion of a bone radiograph. 4. The computer learns which shapes and objects can be used to identify a sarcoma in a human bone radiograph