Let us now give some details on the IvisionLAB dataset composition, and show the performances our approach can achieve on the test dataset.

The DNS Panoramic images from IvisionLAB [2] were used to train our Modified U-Net network. This dataset contains $543$ images of size $1127\times 1191$ annotated with the tooth numbering (FDI notation, i.e. position label) using the Coco format. The images can be divided into 8 categories regarding the presence or absence of all teeth in images, of restorations and appliances (interested readers might refer to [2] for a detailed description of different categories).

Mask R-CNN and U-Net models were trained using a 4-fold cross validation. In total, $111$ images were retained to build the test set and the rest of the images were divided into 4 folds ($108$ images each) thus composing the train and validation data in a cross-validation fashion.

Fig. 2 shows the overall results of the proposed method in terms of average Dice coefficient index (%) for each tooth position from the test dataset.

Using the test dataset, we observe a similar behaviour of the cross validation procedure presented in the paper. The Modified U-Net configuration consistently outperforms the original U-Net model for all teeth classes. The Optimal U-Net offered the best performances with an average dice score of $94.49\%$ followed by the Modified U-Net with an average Dice of $90\%$. Unsurprisingly, the U-Net model exhibits the worst Dice coefficients with an average of $86\%$. The quite large Dice score difference between the original U-Net and the Modified U-Net comes from the detection inaccuracies of the molar teeth. This stems from the fact that $7.9\%$ of molars were missing in our test set (composed of $111$ X-rays) which leads us to question if the misclassification issue on panoramic X-ray images of the original U-Net is not due to missing teeth. So rather than considering tooth classes, it seems interesting to analyse the performances of our model on panoramic X-Rays where some teeth are missing compared to the panoramic images where all teeth are present. For this purpose, we have divided the test dataset into 2 distinct sets: a first one (A) composed of 71 images where all teeth are present, and a second dataset (B), composed of 40 images, in which some teeth are missing. Table 1.2 summarizes the average dice coefficient and its standard deviation for each architecture and with respect to these two sets.

For more quantitative results and comparison of dice coefficients for all tooth classes between the sets A and B, please refer to Fig 3.

The difference of Dice ratio between the Modified U-Net and the Optimal U-Net for set (B) is explained by the rare miss-labeling of teeth by the object detection model. Table 1.2 presents some results of the Object detection Model on the test set and on set (B). From Table 1.2, we can witness 34 miss-classified detections among a total of 36 in the set (B). In fact when a tooth is missing, the model detection network could in some cases classify its nearby tooth by the position of the missing one. Figures 4 and 5 present confusion matrices for a $0.5$ IoU detection threshold on the set (B), corresponding respectively to the upper teeth and to the lower teeth. We observe that the rare miss-classification occur especially for teeth being numbered as one of their missing neighbors.

In order to implement a deep neural network that can automatically detect and segment teeth in panoramic x-ray images, the acquisition of a large dataset with high variability is important and its composition has a decisive impact on the model's performances. Here, we try to evaluate our model's performances on a different dataset that the one we used earlier. So we have collected 81 images from the dentistry department of the University hospital in Nantes. The provided images were then classified into 6 categories according to the presence/absence of all teeth, and the presence/absence of restorations. The images categorization was performed manually, individually selecting the images and verifying the number and the characteristics of the teeth (See Table 2.1).

The second step consisted of annotating the images (obtaining the multi-channel binary images), which corresponds to the detouring of each tooth position in the panoramic X-ray. This was accomplished by 4 different dental students of the University of Nantes. Then the images were cut off to disregard non relevant information (white border around the images and bones parts) with the constraint of keeping the object of interests (Teeth). The cropped images were saved in the new dimension of $1135\times 2236$

Finally, using successive morphological operations of dilatation and erosion, we were able to separate overlapping teeth in the ground truth segmentation.

It is worth noting that the ODON data set is a relatively small dataset (less than 100 images) containing X-ray images varying mostly with respect to the number of teeth and with many missing teeth. Moreover it contains images of inhomogeneous qualities (strongly varying noises and contrasts). To summarize, the IvisionLAB dataset is quite homogeneous but rather large, whereas our own is much smaller but with a larger diversity. In order to investigate whether our method generalizes well, it has been tested using the manual bounding boxes onto these two strongly different datasets: Nantes University dataset (containing 81 images), and a subset of IvisionLab dataset (81 images collected from the 543 initial dataset). Note that we deliberately picked up 81 images in order to have the same number of images in both datasets for an exhaustive comparison. Moreover, we also tried to have a similar distribution between the 2 datasets with respect to the number of panoramic x-rays with missing teeth and with restorations.

As the number of images is small, we chose to use manual bounding boxes to train our network. The idea is to test the generalization of the Modified U-Net using 2 different dataset acquisitions. However, it is worth noting that an object detection network (Mask R-CNN) must be trained with a larger dataset for a better precision as the one used in the paper [1].

To train our network, we fixed for both datasets, the training set to have 51 images, the validation set to have 18 images and the test set to have 12 images. As in [1], we use Dice loss as the criterion to optimize the U-Net model parameters. Moreover, the Adam optimizer is used to train our model with an initial learning rate of $0.0001$. We set the number of training epochs to be $100$ with a batch size of 2.

In this section, we present only a brief summary of the results based on the Dice coefficient. Figures 7 and 8 show the overall results of the proposed method in terms of the average Dice coefficient index (%) for each tooth position with respect to the two datasets. For the two totally different datasets, we observe a comparable performance on each tooth class. Using unseen data, the model offered very good performances with an average Dice score of $93.11\%$ for NantesUniv dataset and $93.39\%$ for the portion of the IvisionLab dataset.

As described earlier, in order to automatically recognize the teeth and various dental treatments on panoramic x-rays, we used the combination of 2 neural networks allowing us to segment and classify the teeth in the X-rays. So far, the NantesUniv dataset consists only of 81 images presenting many missing teeth and various dental treatments. Using manual bounding boxes to train our Modified U-Net network showed very good performances on teeth segmentation. However a larger dataset is needed to train an object detection network for a high detection accuracy. While waiting the acquisition of a larger dataset with high variability, using some data augmentation transformations could be helpful.

References