Abstract: The findings demonstrate the potential of our approach as a robust auto-segmentation system on OCLs in CBCT images, while the segmented masks can be used to further improve OCLs dental diagnostic efficiency.
The Challenge of Jaw Lesion Diagnosis
Odontogenic Cystic Lesions (OCLs) represent a diverse group of pathologies in the jaw, ranging from common benign cysts to aggressive tumors such as ameloblastomas. While Cone-Beam Computed Tomography (CBCT) is the standard imaging modality for these conditions, manual diagnosis remains a significant challenge. Different types of lesions often exhibit remarkable similarities in shape, location, and internal texture, leading to potential misdiagnosis and high clinical workload.
From a computational perspective, two major hurdles complicate automated segmentation: the extreme variability in lesion scale and the high background-to-lesion ratio. Standard neural networks often struggle to distinguish small early-stage cysts from the surrounding healthy bone and teeth in a full-head CBCT scan.
Methodology: The OCL-Net Framework
To address these challenges, researchers developed OCL-Net, a deep learning framework built upon the 3D U-Net architecture, incorporating three novel enhancements:
- Auto-Adapting Schema: This two-stage training process allows the model to focus on relevant anatomy. After an initial training phase, the model generates a preliminary mask to automatically crop the input images around the lesion. This "zoom-in" effect eliminates redundant background data, significantly improving the model's focus on pathological features.
- Multi-scale Dense Attention Module: To manage size variability, the architecture integrates attention maps across four different scales of the decoder. This mimics the clinical workflow where radiologists first identify a region of interest at a coarse scale before examining fine-grained border textures.
- Size-aware Dice Loss: Standard loss functions can be unstable when segmenting very small objects. OCL-Net utilizes a size-aware component that penalizes the model based on the ratio of predicted size versus ground truth, ensuring robust performance across all lesion volumes.
Experimental Results
The study utilized a curated dataset of 300 anonymous CBCT images from the Stomatological Hospital of Wuhan University, including Radicular Cysts (RC), Dentigerous Cysts (DC), Odontogenic Keratocysts (OKC), and Ameloblastomas (AM). Ground truth labels were manually annotated by three radiological experts.
OCL-Net achieved superior performance compared to standard baselines, reaching an overall Dice score of 88.84%, an IoU of 81.23%, and an AUC of 92.37%. Ablation studies confirmed that the "Auto-Adapting" mechanism alone improved the Dice score by over 1% by discarding irrelevant jawbone information.
Clinical Impact and Future Directions
Beyond technical metrics, the study evaluated the system's utility in a real-world clinical setting. In tests involving seven clinicians, the use of OCL-Net-generated masks led to a modest increase in diagnostic accuracy (from ~53% to ~56%). More significantly, the time required for diagnosis dropped from 101 seconds to 47 seconds per case (P < 0.05). This suggests that the primary value of AI in this context may be as a "speed multiplier," reducing clinician burnout and increasing patient throughput.
Despite these successes, the Odontogenic Keratocyst (OKC) remains a diagnostic challenge, often being confused with ameloblastomas due to their similar radiographic appearance. Future work will focus on expanding the dataset to include more diverse lesion subtypes and developing downstream tasks to automatically predict pathological types from the segmented volumes.
Summary updated