[Paper Review] Detection of Tooth caries in Bitewing Radiographs using Deep Learning

Muktabh Mayank Srivastava, Pratyush Kumar+2 more|arXiv (Cornell University)|Nov 20, 2017

Dental Radiography and Imaging3 references61 citations

TL;DR

The paper presents a deep FCNN-based CAD system to automatically detect dental caries in bitewing radiographs, achieving higher recall and F1-score than three practising dentists on a test set.

ABSTRACT

We develop a Computer Aided Diagnosis (CAD) system, which enhances the performance of dentists in detecting wide range of dental caries. The CAD System achieves this by acting as a second opinion for the dentists with way higher sensitivity on the task of detecting cavities than the dentists themselves. We develop annotated dataset of more than 3000 bitewing radiographs and utilize it for developing a system for automated diagnosis of dental caries. Our system consists of a deep fully convolutional neural network (FCNN) consisting 100+ layers, which is trained to mark caries on bitewing radiographs. We have compared the performance of our proposed system with three certified dentists for marking dental caries. We exceed the average performance of the dentists in both recall (sensitivity) and F1-Score (agreement with truth) by a very large margin. Working example of our system is shown in Figure 1.

Motivation & Objective

• Develop a large annotated dataset of bitewing radiographs with clinically verified caries annotations.
• Build an end-to-end deep learning system to localize caries directly from radiographs without specialized preprocessing.
• Evaluate the system against practicing dentists using ground-truth annotations.
• Demonstrate higher sensitivity and competitive agreement with truth compared to dentists.

Proposed method

• Train a 100+ layer Fully Convolutional Neural Network (FCNN) to perform dense per-pixel caries classification on bitewing radiographs.
• Annotations are loose polygons around caries; training treats the task as dense pixel-wise labeling (0/1).
• Create tight bounding boxes around FCNN outputs to form final caries detections.
• Evaluate using IoU/Jaccard with ground-truth polygons, requiring IoU > 0.8 for a successful cavity search.
• Use 5% of the training data as a validation set; no extensive hyperparameter search performed.

Experimental results

Research questions

• RQ1 Can a dense FCNN model trained on annotated bitewing radiographs achieve higher detection sensitivity for dental caries than practicing dentists?
• RQ2 Does automated CAD output provide better agreement with ground truth (F1) than clinicians on a held-out test set?
• RQ3 Is a dense per-pixel approach effective for marking caries given loose polygon annotations?
• RQ4 What is the trade-off between recall and precision for the proposed system compared to clinicians?

Key findings

• The system achieves higher recall (sensitivity) than the three dentists on the test set (System recall 80.5% vs. Dr.1 47.7%, Dr.2 43%, Dr.3 34.4%).
• The system shows competitive to higher precision (61.5%) relative to dentists (Dr.1 63%, Dr.2 81.5%, Dr.3 89.1%).
• The system achieves an F1-score of 70, surpassing all three dentists (Dr.1 54, Dr.2 56, Dr.3 50).