Karianne Ervik Winsnes

• Torgersen, Gerald; Winsnes, Karianne Ervik; Handal, Trude; Diep, My Tien; Ørstavik, Dag Solmund & Sunde, Pia Titterud (2026). Open‐Weight Deep Learning for Periapical Index Scoring: Training and Comparison of Convolutional Neural Network Models Based on Three Distinct Architectures and an Ensemble Approach. International Endodontic Journal. ISSN 0143-2885. doi: 10.1111/iej.70268. Fulltekst i vitenarkiv Vis sammendrag

  Aim This study aimed to (i) train and compare three convolutional neural network (CNN) architectures for Periapical Index (PAI) scoring on intraoral radiographs, and (ii) assess whether an ensemble of these models improves performance compared with any single architecture. Methods A dataset of 17,549 apex‐centred image clips extracted from intraoral radiographs was annotated by a calibrated observer (KW) with PAI scores (1–5) and used to train three CNN architectures (ResNet50, EfficientNet‐B3 and ConvNeXt‐Tiny) for PAI scoring. The image clips were split 80/20 into training and validation sets, and model selection was based on validation performance of quadratic weighted kappa (QWK). The three CNN models were combined into an equal‐weight soft‐voting ensemble model, and all four models were evaluated on an independent balanced test set of 200 images (40 per PAI score). QWK was used as the primary outcome metric, and bootstrapping was used to estimate the confidence intervals (CIs). The PAI scale was dichotomized into healthy (PAI 1–2) and diseased (PAI 3–5), and accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and F1 score were calculated. Results The ensemble model achieved the highest performance with a QWK of 0.79 on the test set, followed by EfficientNet‐B3 (QWK = 0.77), ResNet50 (QWK = 0.76) and ConvNeXt‐Tiny (QWK = 0.70), with overlapping confidence intervals (CIs). For the binary PAI classification, the performance of the three CNN models and the ensemble model was comparable. The ensemble model achieved the following results: accuracy 86%, sensitivity 83%, specificity 90%, PPV 93%, NPV 78% and F1 score 0.88. Conclusion Optimization and evaluation of three CNN models and an ensemble model for PAI scoring showed comparable performance on both the five‐class PAI scale and the binary PAI scale on an independent test set. The ensemble model did not show statistically significant improvement over the three individual CNN architectures. The AI models trained and tested in this study are made open‐source under the open‐source MIT licence.

• Wigen, Tove Irene; Winsnes, Karianne Ervik; Varatharajan, Rahavy & Wang, Nina Johanne (2019). Helsesøstres erfaring med kariesforebyggende arbeid hos barn under 3 år. Den norske tannlegeforenings Tidende. ISSN 0029-2303. 129(5), s. 432–440. Fulltekst i vitenarkiv

Se alle arbeider i NVA

• Torgersen, Gerald R.; Winsnes, Karianne Ervik; Handal, Trude & Sunde, Pia Titterud (2025). Visualizing AI Decisions in Automated Periapical Index Assessment. Fulltekst i vitenarkiv Vis sammendrag

  Objective: Explainable artificial intelligence (XAI) is critical for clinical acceptance of AI models in dental radiology. This study explores the use of XAI—primarily Grad-CAM and LayerCAM—to interpret deep learning-based Periapical Index (PAI) scoring from intraoral radiographs, with a focus on exploring attentions patterns, meaning the areas contributing to the models decision of PAI score. Methods: We trained an open-source deep learning model for PAI classification on 300×300 pixel radiographic clips, using expert-labeled data. Grad-CAM and LayerCAM were applied to visualize model attention, generating class-specific average heatmaps and analyzing individual cases for anatomical plausibility and potential artifacts. Additional XAI methods were explored for comparison. Code and weights will be published together with the main publication. Results: XAI visualizations revealed that model attention patterns systematically varied with PAI class: lower scores showed focused periapical activation, while higher scores exhibited more diffuse and multi-focal attention. Individual heatmaps highlighted both expected anatomical focus and occasional peripheral activations, suggesting areas for further model refinement. LayerCAM and Grad-CAM provided complementary insights into model interpretability. Conclusions: Grad-CAM and related XAI methods offer valuable insights into the decision-making of AI models for PAI scoring. Ongoing work aims to refine XAI techniques and analyze individual attention maps to improve both model performance and interpretability.

• Torgersen, Gerald & Winsnes, Karianne Ervik (2024). Hot Insights: Explainable AI and Heatmaps for Accurate PAI-Scoring in Endodontic Treatments. Fulltekst i vitenarkiv Vis sammendrag

  Introduction The application of Convolutional Neural Networks (CNN) for pathology detection within dental imaging is rapidly expanding(1). Evaluating the outcome of root canal treatments, particularly using the periapical index (PAI)(2), is crucial in endodontics. Traditionally, PAI scoring involves qualitative judgment following a calibration process. However, as Pythagoras famously noted, there is no such thing as quality in the world; everything is based on numbers. Previous studies have demonstrated the potential of CNNs in automating PAI scoring(3,4). Building on this work, our project, in collaboration with endodontic specialist residents, aims to develop a CNN model for accurate PAI scoring on dental radiographs. This year, our focus shifts to enhancing interpretability by visualizing regions of interest using a Gradient-weighted Class Activation Map (Grad-CAM) (5). By doing so, we aim to uncover the underlying quantitative factors that influence PAI classification. This method not only enhances model transparency but also aligns with the principles of explainable AI, providing valuable insights for clinical decision-making. Methods A set of 300 x 300 pixel clips, centered on the apex and scored by an experienced general dental practitioner working at the Department of endodontics, were used for transfer learning on an EfficientNet B3 model(6) pretrained on the ImageNet dataset(7). When inferencing test images we used the Grad-Cam framework to visualize impact of image areas on choosing different classes. Results The heatmaps show attention localized to the apex on the inferred class. Conclusion Heatmaps provide insight into the areas the model focuses on when determining the class. Further analysis will be conducted to utilize this information to enhance the machine learning approach for PAI scoring. We also plan to use the heatmaps as feedback to the human scorers. References 1. Pauwels R. A brief introduction to concepts and applications of artificial intelligence in dental imaging. Oral Radiol. 2021 Jan;37(1):153–60. 2. Ørstavik D, Kerekes K, Eriksen HM. The periapical index: A scoring system for radiographic assessment of apical periodontitis. Dent Traumatol. 1986 Feb;2(1):20–34. 3. Moidu NP, Sharma S, Chawla A, Kumar V, Logani A. Deep learning for categorization of endodontic lesion based on radiographic periapical index scoring system. Clin Oral Investig. 2022 Jan;26(1):651–8. 4. Pauwels R, Brasil DM, Yamasaki MC, Jacobs R, Bosmans H, Freitas DQ, et al. Artificial intelligence for detection of periapical lesions on intraoral radiographs: Comparison between convolutional neural networks and human observers. Oral Surg Oral Med Oral Pathol Oral Radiol. 2021 May;131(5):610–6. 5. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. Int J Comput Vis. 2019 Oct;128(2):336–59. 6. Tan M, Le QV. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. 2019 [cited 2024 Aug 1]; Available from: https://arxiv.org/abs/1905.11946 7. Yang K, Qinami K, Fei-Fei L, Deng J, Russakovsky O. Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchy. In: Conference on Fairness, Accountability, and Transparency. 2020.

• Torgersen, Gerald R. & Winsnes, Karianne Ervik (2024). PAI meets AI – Can Artificial Intelligence Perform Classification of Endodontic Treatment Outcomes? Fulltekst i vitenarkiv Vis sammendrag

  Aim: To identify the most effective convolutional neural network (CNN) model for fine-tuning on a dataset of intraoral radiographs and to evaluate the accuracy of the fine-tuned model in classifying the periapical index (PAI), thereby determining the model‘s capability in accurately reflecting endodontic treatment outcomes. Material and Methods: The model employed anonymized intraoral radiographs with landmarks and PAI scores from endodontic treatments performed at the University of Oslo. Scores were done by master students, residents in endodontics and PHD-candidates. A plugin to ImageJ was used to manually mark landmarks and set the PAI as an indicator for treatment outcome. The PyTorch library in the Python programming language was used to fine-tune and test several pretrained opensource CNN models. Several models including RESNET and VGG19 were fine-tuned on our dataset and tested for accuracy. Results: A fine-tuned VGG19 based model performed best, achieving an accuracy rate of 64%. Conclusion: While applying machine learning to image analysis in standard datasets, such as handwritten digits (MNIST) or flower classification, appears straightforward, the development of a model tailored to endodontic treatment evaluation poses significant challenges. Further results from refinement of training data and experimenting with different CNN models will be presented.

Se alle arbeider i NVA