Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: Study investigating weakly supervised deep learning methods for predicting malignant transformation in oral epithelial dysplasia (OED) using whole-slide histology images (WSIs). A cohort of 163 WSIs (137 OED cases, 50 with malignant transformation) was analysed. A weakly supervised pipeline using IDaRS (Iterative Draw-and-Rank Sampling) with ResNet-34 achieved the highest performance (AUROC = 0.78; F1 = 0.69). Hotspot analysis identified peri-epithelial lymphocyte (PEL) count, epithelium layer nuclei count, and basal layer nuclei count as significant predictors. Survival analysis showed that PELs and epithelial/basal layer nuclear features improve prognostic stratification. The study demonstrates that deep learning can predict malignant transformation and progression-free survival, offering potential support for clinical risk assessment.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study compares automatic deep learning‑based tooth segmentation (DGCNN) with two commercial CAD/CAM segmentation tools (OrthoAnalyzer and Autolign). Using 516 training models and 30 evaluation models, the DGCNN approach achieved high segmentation accuracy and the fastest processing time, demonstrating strong potential for orthodontic diagnostics and appliance manufacturing.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study presents a convolutional neural network (CNN) developed for automatic tooth numbering in panoramic radiographs using the FDI system. A total of 8,000 anonymized images were collected from Asisa Dental S.A.U. centers in Madrid (Spain), curated by two experienced dentists. The model combines Matterport Mask R-CNN for object detection and ResNet-101 for classification, leveraging transfer learning from a previous model that achieved 99.24% accuracy in tooth detection. The network was trained on 1,217 curated images (after filtering and quality control). Training involved 53 runs with 60–300 epochs each, varying learning rates between 0.0014 and 0.012. The final model achieved: Accuracy = 93.83% (total loss 6.17%), Tooth detection = 99.24% accuracy, and Tooth numbering = 93.83% accuracy. It correctly identified missing, filled, and metallic teeth in most clinical cases, though occasional numbering errors occurred for pontics and third molars. The authors conclude that the model is reliable enough for use in clinical environments and demonstrates strong potential for automated diagnostic support.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This preliminary study applies a deep learning object detection approach to contrast-enhanced CT images to automatically detect cervical lymph nodes in patients with oral squamous cell carcinoma. The authors trained a DetectNet model using manually annotated bounding boxes and evaluated performance on an independent test set. The model achieved high precision and moderate recall, with better recall for metastatic nodes at cervical levels IB and II, suggesting potential as a supportive tool to reduce missed nodal findings in busy clinical workflows.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study addresses a key challenge in periodontal diagnostics: accurately estimating the clinical attachment level (CAL) from intraoral radiographs. Although deep learning models can predict CAL bitewing radiographs have a limited field of view preventing CNNs from analyzing anatomical structures that lie outside the captured region. To overcome this limitation the authors developed a generative adversarial inpainting network using partial convolutions to reconstruct missing anatomy and provide additional contextual information for CAL prediction. A large retrospective dataset was used including 80326 images for training 12901 for validation and 10687 for direct comparison between inpainted and non-inpainted methods. Statistical analyses (MBE MAE Dunn’s pairwise test) demonstrated that the inpainting approach significantly improved prediction performance. The MAE decreased from 1.50 mm to 1.04 mm and all pairwise comparisons confirmed superior accuracy for the inpainted models. The study concludes that GAN-based inpainting enhances CAL prediction from bitewing and periapical radiographs and achieves accuracy within the clinically acceptable 1 mm threshold. Clinically this work highlights how AI can compensate for inherent radiographic limitations offering more reliable assessments even when anatomy falls outside the imaging field.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study uses CT images from 66 patients who underwent oral and maxillofacial sur gery (OMS) were landmarked manually in MIMICS. Then the CT slices were exported as images for recreating the 3D volume. The coordinate data of landmarks were further processed in Matlab using a principal component analysis (PCA) method. A patch-based deep neural network model with a three-layer convolutional neural net work (CNN) was trained to obtain landmarks from CT images. The evaluating experiment showed that this CNN model could automatically finish landmarking in an average processing time of 37.871 seconds with an average accuracy of 5.785 mm. It shows a promising potential to relieve the workload of the surgeon and reduces the dependence on human experience for OMS landmarking.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study introduces a novel AI-based pipeline designed not primarily to perform periodontal ultrasound segmentation, but rather to *improve dataset quality* by automatically identifying inadequate or low-quality annotations. Periodontal ultrasonography is an emerging imaging modality capable of visualizing gingival and periodontal soft tissues without ionizing radiation. However, manual segmentation of structures such as the gingival margin, alveolar crest, sulcus, and cemento-enamel junction is highly operator-dependent and prone to inconsistency. These inconsistencies significantly reduce the reliability of datasets used for machine learning model development. The authors developed a deep learning segmentation system using a U‑Net–based architecture trained to detect four periodontal structures. A dataset of 704 ultrasound images was annotated by trained clinicians and divided into 80% training and 20% testing. Importantly, two training strategies were evaluated: (1) a model trained on all annotations, including low-quality ones, and (2) a model trained only on high-quality, curated annotations. The objective was to investigate whether AI could help identify erroneous annotations by exhibiting poor segmentation performance on those cases. The model trained on curated annotations performed substantially better, with Dice scores improving across all four structures. Visual inspection of segmentation failures revealed that the model consistently struggled on images with noisy or incorrect human labels—allowing the system to act as an automated “annotation auditor.” This demonstrates a new conceptual approach: using AI not only for segmentation, but also as a quality control mechanism to enhance dataset reliability before downstream model training. By establishing a quantifiable and automated method for detecting weak annotations, the study provides an important contribution to the emerging field of periodontal ultrasound AI. It highlights the need for rigorous dataset curation and suggests that machine learning can play a central role in improving annotation consistency and accelerating clinical translation of periodontal ultrasound imaging.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study proposes an unsupervised geometric deep learning framework to generate patient‑specific reference bony shape models for orthognathic surgical planning. The method introduces a Surface Deformation Network (SDNet) operating on 3D point clouds to deform a patient’s abnormal cranio‑maxillofacial bone toward a dictionary of normal bony shapes. Sparse representation learning is then used to estimate an objective reference bone model without relying on paired pre‑ and post‑operative data or extensive landmark digitization. The approach significantly improves accuracy compared with landmark‑based sparse representation while preserving midface anatomy.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study introduces AggregateNet, an innovative deep learning model designed for the automated classification of cervical vertebrae maturation (CVM) stages using lateral cephalometric radiographs. A total of 1,018 lateral cephalometrics radiographs were labeled by experts according to CVM stages. The cervical vertebrae were automatically cropped using an object detector, and the model was trained with two inputs:nthe cropped vertebral images, and the patient’s age, which was concatenated with the extracted image feature vector. AggregateNet employs a parallel-structured CNN architecture combined with a pre-processing layer of tunable directional edge-enhancement filters, designed to emphasize morphological contours relevant to CVM assessment. Data augmentation was applied to reduce overfitting, especially because the dataset was separated by gender for improved model fitting. The model’s performance was compared with several well-known architectures—ResNet20, Xception, MobileNetV2, and a custom CNN with directional filters. AggregateNet achieved the highest validation accuracy, reaching 82.35% for female subjects and 75.0% for male subjects. Removing the directional filters leads to a clear drop in performance, highlighting their importance. Overall, the study demonstrates that AggregateNet, combined with directional edge filters, provides superior accuracy for fully automated CVM stage classification, representing a meaningful advancement for AI applications in orthodontics and growth assessment. DOI: https://doi.org/10.1111/ocr.12644

Authors: Dr.Nielsen Santos Pereira

Year: 2026

Description: This study aimed to develop and evaluate an artificial intelligence model using the PyTorch U-Net architecture for the automatic segmentation of the masseter muscle on ultrasonography images. The study utilized a retrospective dataset of 388 images divided into training verification and test sets and the annotations were verified by Oral and Maxillofacial Radiology experts. The model achieved perfect performance metrics (F1 sensitivity and precision of 1.0) suggesting that this deep learning strategy can assist medical practitioners by reducing diagnostic time.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study developed and validated a deep learning–based semantic segmentation model to automatically identify the temporomandibular joint (TMJ) disc on magnetic resonance images. Using a modified U-Net architecture, the model was trained on annotated MR images and evaluated with both internal and external datasets, demonstrating robust performance and clinical applicability despite inter-institutional variability.

Authors: Dr.Nielsen Santos Pereira

Year: 2025

Description: This study proposes a practical deep learning model to predict the time required for mandibular third molar extraction by combining panoramic radiographic images and clinical data. A concatenation approach integrating a Convolutional Neural Network (CNN) and a Multilayer Perceptron (MLP) was developed to simulate clinical decision-making, in which both radiographic appearance and patient-related factors are considered. Using 724 panoramic radiographs and associated clinical variables, the combined model demonstrated strong correlation with actual extraction times, outperforming models based on imaging or clinical data alone. The results suggest that multimodal AI models may support surgical planning, especially for less experienced clinicians. DOI: https://doi.org/10.1186/s12903-022-02614-3