Automatic Lesion Detection in Periapical X-rays

Abstract

In the dental domain, the periapical x-rays play a key role in finding the tooth lesions. However, it's a challenging for a dentist to correctly predict the existence of lesion as well as its breed using the radiographs due to opaque disposition of x-rays. This work aims at developing an application which can automatically detect the type of lesion in periapical x-rays. Two different experiments were performed for this work. In first experiment, the features were extracted using the pretrained network (Alexnet), and then on these extracted features the conventional classifiers Support Vector Machine (SVM) and K-Nearest Neighbor (K-NN) were trained. In Second experiment, the pretrained Alexnet model was fine-tuned on training data. This fine-tuned network was used as feature extractor as well as used to classify the test images using SoftMax function. The seventh Fully Connected (FC) layer of this retained model was used as output for extracted features. On these extracted features the SVM and K-NN classifiers were trained. This second experiment was carried out in two ways. In first way, the pretrained Alexnet model was fine-tuned with augmented data, and in second way, the pretrained Alexnet network was fine-tuned without using the data augmentation technique. The classification accuracy was used to evaluate the performance of both approaches. The highest classification accuracy achieved was 98%.