![References
Add Text
[1] M. M. Lakshmi and P. Chitra, ”Classification of Dental Cavities from X-ray images
using Deep CNN algorithm,” 2020 4th International Conference on Trends in Electronics
and In-formatics (ICOEI), 2020.
[2]. J.H. Lee, D.H. Kim, S.N. Jeong and S.H. Choi, ”Diagnosis and prediction of period on
tally compromised teeth using a deep learning-based convolutional neural network
algorithm” ,Journal of periodontal implant science, vol. 48, no. 2, pp. 114-123, 2018.
[3]. A.A. Al Kheraif, A.A. Wahba and H. Fouad, ”Detection of dental diseases from
radiographic2d dental image using hybrid graph-cut technique and convolutional neural
network”, Mea-surement, vol. 146, pp. 333-342, 2019.
[4]. Oralhealth, https://www.who.int/news-room/fact-sheets/detail/oral-health. Last
accessed 19 Aug 2021.
[5]. J.-H. Lee, D.-H. Kim, S.-N. Jeong, and S.-H. Choi, “Detection and diagnosis of dental
caries using a deep learning-based convolutional neural network algorithm,” Journal of
dentistry, vol.77, pp. 106–111, 2018.](https://image.slidesharecdn.com/midas-220111031151/85/Object-Detection-on-Dental-X-ray-Images-using-R-CNN-24-320.jpg)
Object Detection on Dental X-ray Images using R-CNN
- 1. Object Detection on Dental X-ray Images using Region Based Convolutional Neural Networks Rakib Hossen, Minhazul Arefin and Mohammed Nasir Uddin INTERNATIONAL CONFERENCE ON MACHINE INTELLIGENCE & DATA SCIENCE APPLICATIONS (MIDAS 2021) Date: December 26-27, 2021
- 2. Overview Problem Statement Introduction Objectives Proposed Methodology Result Conclusion Future Work
- 3. Introduction Add Text Object detection is a computer vision technique that allows us to identify and locate objects in an image or video. Tooth decay, also known as dental caries or cavities, is the breakdown of teeth due to acids made by bacteria. Root canal is a dental procedure involving the removal of the soft center of the tooth, the pulp.
- 4. Introduction Add Text In 2020, the WHO estimated that about 50% of the world population is affected by dental caries. In some Asian-Pacific countries, the incidence of oral cancer is within the top 3 of all cancers. Dental caries are cavities or holes (a type of structural damage) in the teeth. Root canal causes due to inflammation or infection in the roots of a tooth.
- 5. Problem Statement Add Text Manual feature extraction for the detection of caries and root canal Difficulties in generating hierarchical features Inaccurate Region Of Interest (ROI) detection and edge detection Difficulties in training model with good convergence
- 6. Objectives Add Text To develop a framework that extracts features efficiently To design an efficient deep learning approach that can aid in the automatic detection of objects in dental X-rays To classify the detected objects in the dental x-ray images into caries and root canal.
- 7. Contributions Add Text Using state-of-the-art model Parameterizing the four co-ordinates of the detected object in the multi scale training Using 12 anchors in the RPN
- 9. Image Acquisition Add Text Source: Digital Dental Periapical X-Ray Dataset Two categories: Caries and Root canal Type: Dental Periapical images 80% for training and 20% for testing
- 10. Model pre-training Add Text Pre-train the model on COCO dataset Fine tune the model Train the model on Digital Dental Periapical X-Ray Dataset Fine tune the model again
- 11. Hard negative mining Add Text Boosts performance especially object detection Hard negatives are the regions where the network has failed to make correct prediction Fed into the network again as a reinforcement for improving trained model Towards fewer false positives and better classification performance
- 12. Number of anchors Add Text Smaller items such as implants seem to be quite frequent in object detection tasks We uses two conditions to assign a positive label to an anchor 𝐿(𝑝𝑖, 𝑡𝑖) = 1 𝑁𝑐𝑙𝑠 𝑖 𝐿𝑐𝑙𝑠 𝑝𝑖, 𝑝𝑖 ∗ + 𝜆 1 𝑁𝑟𝑒𝑔 𝑖 𝑝𝑖 ∗ 𝐿𝑟𝑒𝑔 𝑡𝑖, 𝑡𝑖 ∗ A mini-batch has an anchor 𝑖, and 𝑝𝑖 is a projected probability that anchor 𝑖 will be a real item
- 13. Number of anchors Add Text Most Crucial Part in RPN Traditional Faster R-CNN uses 9 anchors Sometimes fails to detect smaller object In this study we used 12 anchors
- 14. Multi-scale training Add Text Resizes the images to a random scale The detector will be able to learn features across a wide range of sizes Improves the performance towards scale invariance We compute the bounding box regression by parameterized the four co-ordinates of the detected object. It is shown in equation: 𝑡𝑝 = (𝑝 − 𝑝𝑎)/𝑟𝑎, 𝑡𝑞 = (𝑞 − 𝑞𝑎)/𝑠𝑎 𝑡𝑟 = 𝑙𝑜𝑔 𝑟 𝑟𝑎 , 𝑡𝑠 = 𝑙𝑜𝑔 𝑠 𝑠𝑎 𝑡𝑝 ∗ = 𝑝∗ − 𝑝𝑎 𝑟𝑎 , 𝑡𝑞 ∗ = 𝑞∗ − 𝑞𝑎 𝑠𝑎 𝑡𝑟 ∗ = 𝑙𝑜𝑔 𝑟∗ 𝑟𝑎 , 𝑡𝑠 ∗ = 𝑙𝑜𝑔 𝑠∗ 𝑠𝑎
- 15. Feature Concatenation Add Text Feature concatenation is an effective way to add different features together to enhance the classification process. Features are RoI-pooled and L2-normalized from several lower-level convolution layers accordingly. These characteristics are then concatenated and rescaled as if the original scale of the features had not been adopted. A 1×1 convolution is done to match the original network’s number of channels.
- 16. Feature Concatenation Add Text Features from multiple convolutional layers Convolutional layers: lower level & higher level Features: ROI pooled & L2 normalized
- 17. Object Detection Add Text Assign some labels to an object based on their features using soft-max classifier Detect objects in dental x-rays image successfully
- 18. Result Add Text The overlap between two borders is measured by the IoU. The overlap between our anticipated border and the ground reality is then calculated (the real object boundary). Let 𝑛𝑖,𝑗 be the number of pixels of class 𝑖 predicted to belong to class 𝑗, where there are 𝑛𝑐𝑙 different classes, and let 𝑡𝑖 = 𝑗 𝑛𝑖𝑗 be the total number of pixels of class 𝑖. Mean IoU is defined as: 𝑀𝑒𝑎𝑛𝐼𝑜𝑈 = 1 𝑛𝑐𝑙 𝑖 𝑛𝑖𝑖 𝑡𝑖 + 𝑠𝑢𝑚𝑗 𝑛𝑗𝑖 − 𝑛𝑖𝑖
- 19. Result Add Text All ground truth boxes with an IoU ratio less than 0.3 will be labeled as negative Total Loss function per epoc (in thousand)
- 20. Comparison RFCN Resnet 101 68.3% Add Text GoogleNet Inception V3 55.01% Content Here SSD Inception V1 73.56% Add Text Our study 83.45% Content Here Accuracy Error Rate Model
- 21. Comparison
- 22. Conclusion Add Text Successfully detects dental caries and root canal Improvement of Faster R-CNN framework for generic object detection Performs better than other standalone methods Good convergence with better local minima
- 23. Future Works Add Text Will work on huge dataset GPU implementation To build a real time model that can detect dental objects from videos considering the clinical parameters
- 24. References Add Text [1] M. M. Lakshmi and P. Chitra, ”Classification of Dental Cavities from X-ray images using Deep CNN algorithm,” 2020 4th International Conference on Trends in Electronics and In-formatics (ICOEI), 2020. [2]. J.H. Lee, D.H. Kim, S.N. Jeong and S.H. Choi, ”Diagnosis and prediction of period on tally compromised teeth using a deep learning-based convolutional neural network algorithm” ,Journal of periodontal implant science, vol. 48, no. 2, pp. 114-123, 2018. [3]. A.A. Al Kheraif, A.A. Wahba and H. Fouad, ”Detection of dental diseases from radiographic2d dental image using hybrid graph-cut technique and convolutional neural network”, Mea-surement, vol. 146, pp. 333-342, 2019. [4]. Oralhealth, https://www.who.int/news-room/fact-sheets/detail/oral-health. Last accessed 19 Aug 2021. [5]. J.-H. Lee, D.-H. Kim, S.-N. Jeong, and S.-H. Choi, “Detection and diagnosis of dental caries using a deep learning-based convolutional neural network algorithm,” Journal of dentistry, vol.77, pp. 106–111, 2018.
- 25. THANK YOU