Figure 1
From: Generalizability of deep learning models for dental image analysis
Flowchart of the experimental workflow. From both centers, 650 panoramic images were used to train and validate (500 images per center) and test (150 images per center, 100 with apical lesions, 50 without, respectively) models. Image characteristics were compared, the Charité training dataset augmented accordingly and then the re-trained model was tested on KGMU data. Further, models were trained on an increasingly mixed Charité-KGMU dataset and tested on KGMU data. Last, models were tested on subsamples of KGMU data consisting root-canal fillings (and other restorations) or no root-canal fillings/restorations at all.
