Abstract
We aimed to develop an explainable and reliable method to diagnose cysts and tumors of the jaw with massive panoramic radiographs of healthy people based on deep learning. Collecting and labeling massive lesion samples is time-consuming, and existing deep learning-based methods often lack explainability. Based on a collection of 872 lesion samples and 10,000 healthy samples, a two-branch network was proposed for classifying jaw cysts and tumors. The network is first pretrained on massive panoramic radiographs of healthy people and then trained for classifying sample categories and segmenting lesion areas. In the testing stage, 200 healthy samples and 87 lesion samples were included. The average accuracy, precision, sensitivity, specificity, and F1 score for five-class classification were 88.72%, 65.81%, 66.56%, 92.66%, and 66.14%, respectively. For binary classification (lesion vs. healthy), the accuracy reached 90.66%. The proposed method shows encouraging performance and provides a reliable tool for diagnosing jaw tumors and cysts by offering both classification and segmented lesion areas as a diagnostic basis.
Introduction
Odontogenic cysts and tumors of the jaw are often asymptomatic in their early stages, potentially leading to significant bone damage or facial deformity before detection. While routine panoramic radiographs (orthopantomograms) are effective for screening, distinguishing between similar-looking lesions—such as Ameloblastomas (ABs) and Keratocystic Odontogenic Tumors (KCOTs)—remains a challenge even for experienced specialists.
Deep learning (DL) has shown promise in medical imaging but faces two primary hurdles in this domain: data scarcity of rare lesions and the "black box" problem, where models provide diagnoses without visual justification. This study addresses these issues by leveraging a massive dataset of healthy radiographs to establish a baseline for normal anatomy and implementing a dual-branch architecture to enhance explainability.
Methods
The researchers developed a two-branch network designed to improve both diagnostic accuracy and reliability.
Self-Supervised Pretraining
To overcome the limitations of generic pretraining (e.g., on ImageNet), the model's encoder was pretrained using MoCoV2 on 9,500 healthy jaw radiographs. This step allows the model to learn the complex anatomical structures of a healthy jaw, making it more sensitive to pathological abnormalities.
Dual-Branch Architecture and Consistency Constraint
The architecture consists of a shared encoder feeding into two specialized sub-networks: a classification branch and a segmentation branch.
A key innovation is the Position Consistency Constraint ($L_{constrain}$). This loss function forces the classification branch’s attention (visualized via Grad-CAM) to align with the segmentation branch’s predicted mask. This ensures that the model's classification decision is based on the actual lesion area rather than irrelevant image features.
Datasets
The study utilized a comprehensive dataset from the First Affiliated Hospital, Zhejiang University School of Medicine:
- Healthy Samples: 10,000 panoramic radiographs.
- Lesion Samples: 872 total, comprising Dentigerous Cysts (DC, 356), Periapical Cysts (PC, 292), Ameloblastomas (AB, 94), and Keratocystic Odontogenic Tumors (KCOT, 130).
The healthy samples were primarily used for pretraining, while the lesion samples were split into training (70%), validation (20%), and testing (10%) sets.
Results and Discussion
The model demonstrated robust performance across multiple diagnostic tasks. In five-class classification, it achieved an average accuracy of 88.72%. In the binary task of distinguishing healthy tissue from lesions, accuracy rose to 90.66%.
The Value of Healthy Data and Explainability
The inclusion of massive healthy samples as a prior proved critical; domain-specific pretraining led to a 13% increase in accuracy compared to non-specialized models. Furthermore, the dual-branch approach was not merely for visualization; requiring the model to segment the lesion improved classification accuracy by approximately 5%.
As shown in the visual results, the Grad-CAM response maps with the consistency constraint are significantly more focused on the actual lesion areas compared to standard models. This provides a clear "diagnostic basis" that clinicians can trust.
Conclusion
This research presents a reliable and explainable deep learning framework for diagnosing jaw cysts and tumors. By effectively utilizing large-scale healthy data and enforcing spatial consistency between classification and segmentation, the model achieves high diagnostic performance. Future work will focus on integrating attention mechanisms to identify specific biological features and extending this consistency constraint to other medical imaging domains.
Summary updated