To compare these results to a convolutional neural network (CNN), we additionally trained a CNN on our dataset. We chose not to use proprietary commercial software as a benchmark, as its substantially larger training datasets and undisclosed architectures and training procedures preclude reproducible comparisons under conditions comparable to ours. Instead, we adopted the input configuration proposed by 10 , using one image from each anatomical plane (axial, sagittal, and coronal) as input. Given that this architecture achieved competitive results in their study and that we began to observe signs of overfitting, we decided not to include additional slices in the CNN training process. The network consists of two convolutional blocks, each comprising a 3×3 convolution with 32 feature maps, followed by a ReLU activation and a 2×2 max pooling layer. The extracted features are flattened and passed through a fully-connected layer with 128 units and ReLU activation, with dropout applied for regularization, before a final linear layer outputs class probabilities. Input images were normalized and augmented during training with random resized cropping, Gaussian noise, and random erasing. The model was trained using the Adam optimizer with a learning rate of 10−3 for 500 epochs and achieved test accuracies of 70.27% for FDI labeling and 86.67% for diagnostics. Appendix A.6 summarizes the selection of the CNN baseline and the alternatives we considered. Table 2 : Scores reached in the diagnosis task. Labels Accuracy (%) recall precision # of Features Composite yes/no 98.17 0.97 0.95 1811 Impacted yes/no 100 1 1 711 Covered by bone yes/no 100 1 1 711 Attrition/nothing 99.58 0.67 0.67 1211 Filling/Retainer 100 1 1 211 Position of the filling 96.00 - - 464 Secondary caries 95.56 0.8 1 111
Table 3 : Iterated scores for all final predictions, together with overall averages. FDI-Labels Acc. (%) 12, 11, 21, 22 96.60 13, 23 97.83 14, 15, 24, 25 96.60 16, 17, 26, 27 96.60 42, 41, 31, 32 98.57 43, 33 98.57 44, 45, 34, 35 98.57 46, 47, 36, 37 98.57 Diagnostic Labels Acc. (%) recall precision Impacted 98.17 0.97 0.95 Covered by bone 98.17 0.97 0.95 Attrition 97.75 0.67 0.67 No findings 97.75 0.97 0.95 Filling (with pos.) 94.24 0.98 0.94 Secondary caries 88.44 0.8 0.94 Retainer 98.17 0.97 0.95 Our methods (%) CNN reference method (%) FDI-Labels 97.67 70.27 Diagnostic 96.77 86.67
4 Conclusion
We introduce an automated pipeline for tooth classification and diagnosis. Regions of interest (ROIs) are extracted from CBCT data using deep learning methods, while classification is performed via persistent homology. On datasets comparable to those reported in the literature, we reach average accuracy scores of 97.67% for tooth labeling and 96.77% for diagnostic tasks, outperforming a CNN trained on the same data with accuracy of 70.27% and 86.67% respectively.
The labeling and diagnosing in our workflow relies on small datasets and few features, enabling faster and less computationally demanding training. It also accommodates regions of interest of varying size.
By automating tooth classification, our approach reduces clinician workload and supports more efficient diagnostics. Future work includes increasing the sample size to further improve performance and extending the persistent homology framework to tooth detection itself, thereby eliminating the need for a separate segmentation step.
acknowledgments-disclosure-of-funding. This work has received funding from the Swiss State Secretariat for Education, Research, and Innovation (SERI).
Appendix A First Appendix
We will now elaborate on the ablation studies we performed and the decisions we took, in order to settle on A.1 the normalization method, A.2 the tooth detection and its impact on the ROI, A.3 the location of two dimensional slices, A.4 the vectorization method, A.5 the SVM structure, A.6 the choice of CNN. Since the methodology provided in this section is not always the final approach, all scores should be understood relatively within the different experimentation rather than absolutely. We started by using the three-dimensional voxel set of the whole tooth, together with five two-dimensional slices in each anatomical direction. From there we used the filtration as seen in Subsubsection 2.1.2 .