Deep Learning System Diagnoses and Predicts Recurrence of Jaw Tumours
Researchers developed a multimodal AI system for odontogenic keratocyst that fuses histopathology slides with clinical parameters for diagnosis, plus an interpretable model for recurrence risk prediction. The system outperformed existing models in both tasks on collected datasets. Clinical relevance for diagnostics labs is limited to oral pathology, but the multimodal fusion approach is notable.
The original study
A deep learning approach for the diagnosis and recurrence prediction of OKC.
- Authors
- Chen W, Qian M, Zhang M, Yang W, Gao R, Zhang J, et al.
- Journal
- Scientific reports
- PMID
- 41872434
Original abstract
Odontogenic Keratocyst (OKC) is a benign jaw tumor characterized by a high recurrence rate. However, its clinical diagnosis presents significant challenges due to complex pathological morphologies and inconspicuous early symptoms. In this paper, we propose an intelligent OKC diagnosis system that leverages multimodal data fusion and interpretable analysis. Specifically, we first devise a Multimodal Feature-based Diagnosis Model (MFDM), which utilizes a Spatial Feature Fusion Module (SFFM) to fuse features extracted from patients' oral pathological slides and tabular clinical parameters. Second, we formulate an Interpretable Recurrence Prediction Model (IRPM) that extracts features from patient demographics and medical history records. This model harnesses an attention mechanism to weight these features, thereby realizing a quantitative estimation of OKC recurrence risk. Furthermore, we implement an integrated information management platform that deploys both MFDM and IRPM. This platform facilitates OKC detection, risk prediction, case storage, and key data visualization, significantly facilitating the algorithm's usability. We conduct extensive comparative and ablation experiments on collected datasets. The results demonstrate that our approach achieves higher accuracy in both diagnosis and recurrence prediction compared to existing state-of-the-art models. Moreover, the visualization elucidates the rationale behind the model's decision-making process, reinforcing its interpretability and credibility for clinical applications.