| The integration of AI into dental diagnostics is showing promising results in improving accuracy and decision-making. A study was conducted to assess how AI and second opinion frameworks can improve the diagnostic accuracy of dental students. |
| Study Design |
| Participants: |
 |
25 Dentistry Students (22 Female, 3 Male). |
|
4th & 5th Year Students of the University of Copenhagen with Advanced courses in caries pathology & treatment. |
|
|
| AI Training: |
|
The AI model was trained on a machine with a Titan X GPU over 50 epochs. |
|
Data transformations included flipping, rotating, and scaling images, along with adding Gaussian noise. |
|
The dataset consisted of 290 cases from patients aged 18 to 89, with most lesions being approximal (96.2%). |
|
|
| Clinical Findings |
| Treatment Groups: |
|
Patients were divided into two groups: 142 treated with selective removal (SW) and 148 with nonselective excavation (NSE). |
|
Pulp exposure occurred in 19% of the SW group and 29% of the NSE group. |
|
Preoperative pain was reported in 35% of cases, with mild to moderate levels. |
|
|
| Diagnostic Performance |
|
The average F1-score for dental students was 0.586, while the AI system achieved 0.71. |
|
For cases of disagreement with the AI, the students' F1-score dropped to 0.289. |
|
Introducing a second opinion improved the average F1-score for the disagreement subset to 0.468 and raised the overall score to 0.645, particularly benefiting participants with lower baseline performance. |
|
|
| Trust in AI Systems |
|
A key challenge is the trust clinical specialists have in AI-based systems. Errors by the AI can diminish confidence, even among experts. Dental students exhibited distrust despite assurances of the AI’s accuracy. |
|
|
| Addressing the Trust Issue |
|
To mitigate trust issues, the AI was designed as a secondary tool in the diagnostic pipeline, prompting second opinions only in cases of uncertainty. This approach fosters trust and aids decision-making without overtly dictating outcomes. |
|
|
| Simulation Insights |
| Qualitative Case Analysis: |
|
Small factors like radiograph angle or lesion depth can complicate predictions, leading to high misdiagnosis rates in certain cases. |
|
|
| F1-Score Improvement: |
|
The introduction of AI second opinions improved the F1-score from 0.586 to 0.645, enhancing decision-making efficiency, particularly for low-performing participants. |
|
|
| Second Opinion Framework: A Practical Approach |
|
This framework serves as a risk-averse tool that enhances decision-making. The AI requests a second opinion when its prediction differs from the initial diagnosis, allowing for human verification and reducing variability in treatment decisions. |
| |
 |
| |
The Al-triggered second opinion framework. The input for this framework consists of a preoperative radiograph collected from a patient with advanced caries (A) Dentist 1 evaluates the case, determining whether the scheduled excavation treatment will result in pulp exposure or not (B) In the background of the framework, the Al system (C) also predicts the risk of pulp exposure using the preoperative radiograph (A) and clinical information (E) The prediction of the firstdentist and Al are then compared by the framework (D) If the predictions do not agree, a second dentist (F) is consulted for their decision. |
|
|
| Clinical Significance |
|
Integrating this second opinion framework into clinical practice can help dentists avoid unnecessary pulp exposure, aligning with less-invasive treatment recommendations. The AI’s role in prompting second opinions refines treatment plans while mitigating human error. |
|
|
| Future Applications |
|
The second opinion framework can also be incorporated into dental school curriculums, fostering critical discussions among students on complex cases. This collaborative platform enables students and dentists to compare diagnoses with AI insights, leading to more consistent and accurate treatment decisions. |
|
|
