From pixels to prognosis: Application and challenges of artificial intelligence in oral disease diagnosis

Prologue to discovery

AI’s journey has been marked by alternating periods of rapid advancement (“AI Springs”) and periods of reduced progress (“AI Winters”). Recent breakthroughs in computing power, the availability of vast datasets, and the development of sophisticated algorithms have fuelled a dramatic resurgence.^[6]

Just as Alan Turing’s 1950 Turing Test helped win World War II, we now stand at the precipice of the 4^th Industrial Revolution, where AI is set to transform healthcare, particularly in dentistry. This contemporary “AI Spring” is driving a fundamental shift, promising to reshape industries and redefine the relationship between humans and machines.^[13]

What is ML?

ML is a subset of AI that empowers systems to learn from data and improve their performance on a specific task without being explicitly programmed with rules for every possible scenario. Adding AI, giving it the power to think like the human brain, is what constitutes ML. It allows systems to learn and improve from experience without explicit programming. In essence, the algorithm learns the relationship between input data and the desired output by following a sequence of steps: Input Data → Analyze Data → Find Patterns → Prediction/Decision → Learns from Feedback.^[3]

Algorithms in oral disease diagnosis

The application of AI to oral disease diagnosis relies on an elaborate toolkit of ML algorithms. Just as a dentist integrates various pieces of information, a patient’s history, clinical observations, and radiographic findings, these algorithms combine different strengths to achieve a comprehensive and accurate diagnosis.

Convolutional neural networks (CNNs)

The function of a CNN can be likened to a ‘master detective’ meticulously examining every detail of a radiographic image or clinical photograph, searching for subtle clues.^[2] Convolutional layers employ learnable filters to extract features from the input data by performing convolution operations. Non-linear activation functions are employed, allowing the model to learn intricate patterns. Pooling layers reduce the size of the feature maps, which helps decrease computational complexity and prevent overfitting by retaining only the most important information. The extracted features are flattened and fed into fully connected layers, where the final classification or regression task is performed [Figure 1].

Close

Figure 1:

Architecture of a convolutional neural network for Image Classification. [Source: Reference 2]

Export to PPT

Support vector machines (SVMs)

SVMs take on the role of a highly efficient “sorter.” SVMs are powerful classification algorithms designed to find the optimal boundary.^[14] CNN extracts image features, and the SVM classifies them. The performance of SVM, however, is sensitive to parameter selection; thus, other models are often preferred^[15,16] [Figure 2].

Close

Figure 2:

A schematic diagram for support vector machine training. [Source: Reference 3] The X & Y axis are a visual representation of the relation.

Export to PPT

Random forest algorithms (RFAs)

If CNNs are the “detectives” and SVMs the “sorters,” then RFAs are the expert “consultants.” They are ensemble learning methods that combine the predictions of multiple decision trees.^[17]

A decision tree is a hierarchical structure that splits data based on specific criteria. RFAs improve individual decision trees through bagging and random subspace techniques. Bagging trains each tree on a random data sample, reducing variance, while the random subspace method selects a subset of features to ensure model diversity. The final prediction is an ensemble output, combining all trees for improved accuracy and robustness^[7,18] [Figure 3].

Close

Figure 3:

Ensemble learning in random forest: aggregation of decision trees for improved prediction. [Source: Reference 7]

Export to PPT

Visual transformers (ViTs)

Unlike CNNs, which focus on local patterns, ViTs can capture global relationships.^[19] Image patching divides an image into smaller segments, with positional embeddings included to retain spatial relationships. The transformer encoder, using self-attention, assigns importance to patches, while multi-layer perceptrons refine the representations. Finally, the classification head processes the output to make the final prediction^[8] [Figure 4].^[20]

Close

Figure 4:

Architecture of the vision transformer for caries classification.[Source: Reference 20]

Export to PPT

Dental caries

CNNs identify subtle radiolucencies indicative of enamel and dentin demineralization, often at earlier stages than are detectable by the unaided human eye.^[21] Studies have consistently shown high sensitivity and specificity for CNN-based caries detection, with some systems achieving performance comparable to, or surpassing, that of experienced dentists. Specialized CNN architectures, such as U-Net, have proven particularly effective for segmenting carious lesions, providing precise delineation of the affected area.^[4] RFAs have been successfully used to predict caries risk by integrating diverse data sources, including radiographic findings, clinical parameters, patient demographics, and behavioral and dietary factors.^[22] Real-time applications, such as “You Only Look Once” (YOLO) family of models, have also demonstrated high precision and recall in detecting interproximal caries, even under complex conditions, including restorations and orthodontic appliances.^[23] The integration of AI in caries detection is advancing from simple identification to comprehensive lesion segmentation, classification, and risk assessment, thus supporting more precise and individualized dental care.

Periodontal disease

CNNs have been applied to accurately quantify bone loss, track changes over time, and assist in classifying the stages of periodontitis, achieving a level of accuracy comparable to trained periodontists. RFAs and SVMs have been utilized to analyze clinical parameters, including probing depths, bleeding on probing, and tooth mobility, to assess periodontal disease status and predict its progression.^[24] Deep learning models such as YOLOv8 and ResNet18 have achieved up to 97% accuracy in detecting radiographic bone loss and furcation involvement.^[25] Vision Transformers have further improved classification tasks, while AI applied to cone beam computed tomography (CBCT) enables 3D analysis of complex defects, aiding in precise and reproducible diagnosis.^[11]

Challenges in AI implementation

Despite the significant promise and progress of AI in oral disease diagnosis, several critical challenges must be addressed to ensure its successful and responsible implementation. These span technical, ethical, and practical domains:

• Data privacy and security: Compliance with regulations such as HIPAA and GDPR is paramount.^[17] These require anonymization, secure storage, and patient consent.^[32]

• Limited datasets and data quality: Deep learning relies on large, high-quality, and diverse datasets. In oral health, datasets are often small, poorly labeled, or lack diversity. Class imbalance is a concern.^[5]

• Regulatory compliance and validation: AI diagnostic tools are medical devices needing regulatory approval (e.g., Food and Drug Administration), requiring rigorous validation for safety, effectiveness, and utility.^[5] Regulatory frameworks are still evolving.

• Interpretability and explainability: Complex deep learning models often operate as “black boxes.” Clinicians need transparent decision-making to support clinical judgment.^[22]

• Cost and accessibility: High development and implementation costs may limit accessibility.^[5]