Computer tomographic differential diagnosis of ameloblastoma and odontogenic keratocyst: classification using a convolutional neural network

Abstract

Objective:

To analyse the automatic classification performance of a convolutional neural network (CNN), Google Inception v3, using tomographic images of odontogenic keratocysts (OKCs) and ameloblastomas (AMs).

Methods:

For construction of the database, we selected axial multidetector CT images from patients with confirmed AM (n = 22) and OKC (n = 18) based on a conclusive histopathological report. The images (n = 350) were segmented manually and data augmentation algorithms were applied, totalling 2500 images. The k-fold × five cross-validation method (k = 2) was used to estimate the accuracy of the CNN model.

Results:

The accuracy and standard deviation (%) of cross-validation for the five iterations performed were 90.16 ± 0.95, 91.37 ± 0.57, 91.62 ± 0.19, 92.48 ± 0.16 and 91.21 ± 0.87, respectively. A higher error rate was observed for the classification of AM images.

Conclusion:

This study demonstrated a high classification accuracy of Google Inception v3 for tomographic images of OKCs and AMs. However, AMs images presented the higher error rate.

Introduction

Intraosseous benign odontogenic lesions comprise a group of pathological entities characterized by variable symptoms, clinical presentations and radiographic and histological features. These lesions are detected incidentally during routine radiographic examination since symptoms are uncommon at early stages. Although benign, these lesions exhibit a destructive potential associated with high rates of recurrence and can reach a large size until they become symptomatic,¹ reducing the quality of life of affected patients. Excluding inflammatory pathologies and odontoma, odontogenic keratocyst (OKC) and ameloblastoma (AM) are the most prevalent benign intraosseous lesions cited in different studies.^1–4 The similar imaging findings of these lesions render their distinction a challenge.⁵

OKC is the third most common odontogenic cyst that frequently affects younger individuals, with a peak incidence between the second and third decades of life.^3,6 Radiographically, OKC appears as a well-defined uni- or multilocular radiolucency associated with minimal buccolingual expansion.^6–8 This odontogenic cyst frequently involves the posterior mandible and tends to grow towards the mandibular ramus.^9,10 Treatment generally consists of enucleation combined with other adjuvant therapies in order to reduce the risk of recurrence.^10,11

AM is the most common epithelial odontogenic tumour,⁴ which frequently affects the posterior mandible. A peak incidence between the fourth and sixth decades of life has been reported.⁸ Radiographically, AM appears as a uni- or multilocular radiolucent lesion associated with impacted teeth, tooth displacement, root resorption, cortical bone expansion and/or perforation, and the presence of septa that confer a soap bubble or honeycomb appearance.^8,12 In view of the high risk of recurrence, surgical resection is usually the therapy of choice for this tumour.¹³

Since different treatment approaches exist for these pathologies and their imaging features may be confused, the subjective interpretation by the examiner to identify particularities is time-consuming and relies on his/her experience for suggesting a diagnosis as there are no well-defined criteria to distinguish these tumors. Within this context, auxiliary tools such as the assessment of intralesional heterogeneity using Hounsfield unit (HU) may contribute to this distinction.¹⁴

In the field of computer science, artificial intelligence is the study of computer algorithms and models that simulate learning processes and problem-solving of humans. An interesting subdivision of this field is convolutional neural networks (CNNs), which consist of algorithms inspired by the structure of the human brain, with a function analogue to that of neuronal cells.¹⁵ It is important to highlight that artificial intelligence is not a new technology, but the advances in the processing capacity of machines in recent years have made its use possible and have increased interest and knowledge in the area.

In medicine, studies evaluating the use of neural networks as an auxiliary diagnostic tool have reported promising results regarding the detection and classification of malignant lesions.^16–19 Although introduced later in dentistry, studies also suggest an excellent performance of CNNs for the assessment of radiographic images, with different applications^20–24 that range from the detection of caries lesions in two-dimensional exams²⁵ to the diagnosis of Sjögren’s syndrome using more complex methods such as CT.²⁶

Considering the similar imaging features of OKCs and AMs and the importance of an accurate presumptive diagnosis for adequate therapy, objective tools can greatly assist in the diagnostic process. Within this context, the application of CNNs is a possibility of combining supervised learning algorithms with imaging methods in an attempt to increase diagnostic accuracy, to reduce the evaluation time and to establish the most appropriate therapeutic approach considering the high recurrence rates of OKCs and AMs.¹¹

Understanding this problem, the aim of this study is to analyse the automatic classification performance of a CNN model using tomographic images of OKCs and AMs.

Methods and materials

This study was conducted at the Dental Radiology Service, School of Dentistry, Federal University of Bahia (UFBA), in collaboration with the Postgraduate Program in Computer Science (PGComp/UFBA) and the Federal Institute of Education, Science and Technology of Bahia (IFBaiano). The project was approved by the Ethics Committee under approval number 366.989.

A total of 48 multidetector computed tomography (MDCT) exams of patients seen at the Oral-Maxillofacial Surgery and Traumatology Service of the School of Dentistry, UFBA, and Santo Antônio Hospital (Salvador, Bahia, Brazil) between 2013 and 2019 were selected. Images with a conclusive anatomopathological report of AM or OKC provided by an experienced oral pathologist were included. Exclusion criteria were as follows: recent intervention at the lesion site, clinically evident infection and lesions > 80 mm. We selected these conditions because they could result in significant changes in content and clinical behaviour. Images with a large number of metallic artefacts involving the region of the lesion were also excluded from the data set, for a total of eight patients excluded.

All images were acquired by HiSpeed CT (GE Healthcare, Chicago, USA) in 0.6 mm with 15.8 FOV, 512 × 512 matrix. The MDCT scans were processed with the soft tissue customized window of the Horos^™ DICOM Viewer software (Pixmeo, SARL, Bernex, Switzerland) in order to identify differences in the density of the intralesional tissue. The tomographic sections comprising the image compatible with the lesion were exported from the DICOM (Digital Imaging and Communications in Medicine) to TIFF format (Tagged Image File Format) and then segmented.

The images were segmented manually and individually into rectangular ROIs by a single experienced and calibrated examiner using the ImageJ software (National Institutes of Health, Bethesda, Maryland, USA), which resulted in arbitrary regions of interest of different sizes as shown in Figure 1. The images were saved in PNG (Portable Network Graphics) format, the standard format of CNN which generates compressed files without loss of information.

Figure 1.

TCMD. Axial view. Soft tissue customized window (WL 76, WW 209); b, c and d show segmented ROIs of axial sections of TCMD obtained from different patients with Am and f, g and h show ROIs after segmentation of axial sections of TCMD obtained from different patients with OKC.

Construction of the dataset

Axial tomographic images of 22 AMs and 18 OKCs were selected for construction of the dataset, resulting in a total of 350 samples of the original set of images. Only sections in which the tumour was visible and measured less than 80 mm were selected.

Data augmentation is usually applied to increase the number of images and to minimize overfitting, that is when a CNN model performs well during training but generates many classification errors when exposed to a new dataset. This technique consists of artificially augmenting a dataset using transformation methods in which new images are produced from the original images.^27,28 For this purpose, rotation, mirroring, random displacement in the horizontal and vertical directions, elastic distortion and resizing were applied to the original images of the dataset. During the augmentation process, not all of these transformations were applied to the images. For each transformation, a probability to its application can be defined. Data augmentation is a common procedure in medical image datasets to deal with the usually smaller number of samples. Previous studies^29,30 analysed the use of such procedures in medical image datasets and concluded that it is an important step to improve the learning process. There are recent works that propose the use of CNNs to produce more parametrized augmentation images in the medical image domain.^31,32 In this study, data augmentation was applied to our 350 images dataset, using the transformations rotation (90^o and 180^o), flips (horizontal and vertical), elastic distortion (configure to be moderate) and resizing (to regularize the different ROIs size). The result was a 7.14 times augmentation, which means 2500 images. All images, original and augmented, were resized to the fixed resolution of 299 × 299 pixels, as required by the CNN Inception v3.

The database was then divided into two main sets: a development set, used to learn the parameters, to estimate the generalization error during or after training, and a final test set used for final validation of the network.³³ For this study, 70% of the samples from the original database were selected for the development set and the last 30% of the samples were selected for the final test set. It is important to notice that these two sets, development and final test, are disjunct, that means images in one set did not appear in the other. This selection was done manually, arranging the images on different directories.

The development set was subdivided into three subsets: training, validation and test, as shown in Figure 2. In this subdivision, 10% of the samples were used to define the test set and the remaining 90% of the samples of the development set were divided equally to build the training and validation sets. It is important to notice that the test set was built once and the same set of samples was used during the training/validation steps (Figure 2), related to the cross-validation process, described afterwards.

Figure 2.

Schematic representation of the 2-fold ×5 method used to promote the cross-validation procedure during the learning phase. The development data consist of 70% of the original dataset and is selected manually. The test set corresponds to 10% of the development dataset.

Cross-validation

One method used for evaluating the generalization capability of a model is the cross-validation. It is a statistical technique that can determine the robustness of a model.³⁴ The basic concept of this method is the subdivision of the dataset into mutual exclusive subsets, and then use some of these subsets to estimate the model parameters (training), while the other subsets are used to validate the model.³⁵ There are many ways to produce such partitioning in the literature. In this research, we used the k-fold cross-validation, which subdivides the training set into k folds with nearly equally size subsets or folds, containing samples distributed randomly. Then, k iterations of training and validation steps are applied using a different fold of data for validation and the remaining k-1 folds for learning/training. Special care should be taken to ensure that each fold is statistically representative of the original dataset.³⁶

More precisely, this study used the 2-fold ×5 scheme, proposed by Dietterich.³⁷ Figure 2 depicts this scheme: the development set was divided into 2-folds that are exchanged during the cross-validation process. The cross-validation is repeated five times. Of course, the folds are generated from scratch for each one of the five cross-validation iterations. The test dataset is kept unchanged during this process. According to Dietterich,³⁷ this scheme improves the test stability and reduces the computational cost. Also, in the context of this study, where we aim to show that CNNs can improve the differentiation of two kinds of lesions, we believe that this approach was appropriate.

Convolutional neural network architecture

The CNNs are one of the most advanced deep learning architectures.³⁸ This learning approach allows computational models composed of multiple processing layers to learn data representations with multiple levels of abstraction. In this study, we used the Inception v3 model developed by Google LLC (Mountain View, CA, USA) (Figure 3).

Figure 3.

Schematic representation of the Google Inception v3 model architecture.

The typical architecture of a CNN comprises a series of stages. The first stage consists of two layers responsible for feature extraction: convolutional layers, which process inputs considering local receptive fields, and the pooling layer, which reduces the spatial dimensionality of the representations. The second stage consists of a fully connected layer, which acts as the classifier itself. The output of the CNN defines the probability that the image belongs to one of the classes for which the network was trained.³⁹

Inception model

The architecture of the Inception v3 model consists of a stack of modules that look like small independent networks, divided into several parallel branches⁴⁰ and is designed to extract features with a reduced number of images, thus decreasing the computational cost of accurate classification of images using the deep learning method. The aim of this model, which was introduced by Szegedy,⁴¹ is to act as a multi-level feature extractor, executing convolutions with three different kernel sizes (1 × 1, 3 × 3 and 5 × 5) and thus covering elements of interest with different sizes and positions in an image. The Inception model was run applying the RMSprop optimizer, which is effective and practical for the learning method used in this study.

All the experiments described in this study were performed on a Dell Inspiron i17-7500U notebook, with an Intel i7 2.7 GHz processor, 8 GB of RAM, NVidia GeForce MX150 GPU, running Linux/Ubuntu 20.04 operating system. The application was written in Python^35,40,42 programming language and used the following packages: TensorFlow v.1.9.0 and Keras v.2.2.0 for the CNN; and Augmentor for the image augmentation process. Although the tests were made in a Linux environment, the software developed is multi-platform and can run on other operating systems such as Windows and MacOS, with minor adjustments, as well. In this machine, the mean time per training was about 42.69 h (without GPU acceleration enabled).

Results

In order to evaluate the learning process, the precision and accuracy of the model were analysed.³⁵ At the end of the 2-fold ×5 cross-validation process, the Google Inception v3 neural network model used for the automatic classification of tomographic images of OKCs and AM reports the precision value and a 2 × 2 matrix, called confusion matrix. Table 1 shows the matrixes for each 2-fold cross-validation step in the five interactions. As our application deals with a 2-class classifier, the values on this matrix correspond to the percentage of true-positives in the main diagonal and the percentage of false-positives and false-negatives in the secondary diagonal. Based on this matrix, many measures can be defined, such as accuracy, precision, sensibility and F1-Score.^35,43 In our analysis, accuracy is the most important measure, indicating the proportion of samples that produces the right classification.

Table 1.

Confusion matrix for the classification of odontogenic keratocysts and ameloblastomas images after data augmentation (n = 2500) in each training run (×5 iterations)

Input	first		second		third		fourth		fifth
	AM	OKC	AM	OKC	AM	OKC	AM	OKC	AM	OKC
AM	0.85	0.15	0.85	0.15	0.84	0.16	0.77	0.23	0.84	0.16
OKC	0.05	0.95	0.17	0.83	0.07	0.93	0.06	0.94	0.09	0.91

AM, Ameloblastoma;OKC, Odontogenic keratocyst.

Mean time per training run: 42.69 h.

In general, the confusion matrix presented in Table 1 indicates a higher error rate for AM images, suggesting a better classification capacity for axial sections of OKCs.

Table 2 shows the accuracy achieved by the CNN in the validation of each iteration. The values exceeded 90% in all five iterations, with a standard deviation less than 1%, suggesting an excellent classification capacity when new images were presented to the tested model.

Table 2.

Accuracy (%) and standard deviation of cross-validation for five iterations

Accuracy	first	second	third	fourth	fifth
Mean ± standard deviation (%)	90.16 ± 0.95	91.37 ± 0.57	91.26 ± 0.19	92.48 ± 0.16	91.21 ± 0.87

Discussion

The present study demonstrated the capacity of the Google Inception v3 CNN model to classify MDCT images of OKCs and AMs with accuracies higher than 90%. This result agrees with findings reported in dental^20,25 and medical studies.^44–46 To our knowledge, this is the first study to test the classification performance for two odontogenic tumours with similar characteristics rather than only the identification of pathological entities in imaging exams.

The data of Table 1 generated by cross-validation repetitions show a higher error rate for the classification of AM images, corroborating the findings of Kwon et al. (2020).⁴⁷ This higher rate might be explained by the radiographic features of this tumor. As suggested by Kitisubkanchana et al. (2020),⁵ OKCs are generally unilocular with smooth borders and exhibit anteroposterior growth without significant cortical bone expansion, while most AMs are multilocular and have irregular borders. This disformed feature of AM, as indicated in Figure 1, may have impaired the recognition of the tomographic image patterns of this tumor by CNN, resulting in classification errors. Furthermore, the existence of multiple fluid-containing cystic areas detected in different regions of AM results in tomographic sections with areas of lower tomographic density,¹⁴ which can confuse the perception of the CNN and increase the probability of errors.

Regarding the content of the two lesions evaluated, it is known that, although not statistically significant, a difference exists between the HU values provided by MDCT.¹⁴ OKCs are composed of poorly soluble proteins such as parakeratin and collagen, which result in lower average attenuation values (HU) compared to AMs. The latter tumor consists of epithelial cells and dense connective tissue that confers higher average attenuation.^14,48 MDCT images were chosen because this method allows differentiating the tissues comprising the tumor mass and content of the cystic cavity by selecting an adequate window for this purpose.⁴⁹

Comparison of the diagnostic accuracy of panoramic radiography and CT for the differentiation of OKCs and AMs showed that the three-dimensional method contributes substantially to the correct diagnosis, particularly because of its capacity to identify high-density areas, number of loculations, and bone expansion.⁴⁸ In the present study, these characteristics may have been detected during the step of feature extraction and used by the CNN for the distinction of the two pathological entities. Some authors have reported promising results regarding the computer-assisted diagnosis using CT images of periapical cysts and OKCs,⁵⁰ radiolucent lesions^51,52 and maxillofacial cysts.^53,54

We highlight that this is the first study evaluating the use of CNNs for the automatic classification of benign odontogenic lesions using MDCT images. Moreover, unlike the present study which evaluated the differentiation of two complex and similar pathological entities, most studies applying CNNs to radiographic images in dentistry used two-dimensional imaging methods and aimed to detect or diagnose the presence of diseases compared to images of healthy structures, therefore with a lower level of complexity than the study described in this article.

Although the diagnostic value of CNN models is considerably high, their application is still not a reality⁵⁵ since the steps of segmentation and training are time-consuming and incur high computational costs, which turns it into a limitation for its usage. The exclusion of large diameter lesions and infected lesions is also a limitation of this study since these characteristics can raise doubts about diagnosis and imitate malignant lesions. In future studies, by including the study of these images, CNN’s performance limits can be better explored. Afterwards one of main limitations of the study is a small sample size. Although the same image used in the training set was not reused for test set, there is a chance that images from the same patient could be used for both processes. Further studies with larger samples should be evaluated to ensure the accuracy and to prevent overfitting.

After successive studies and improvements in the methodological process, these systems should become an integral and essential part of the radiology workflow. The model tested here is intended to be readily available for scientific purpose. It is expected that in the medium term and after expanding the database and new training and test runs in order to ensure high rates of accuracy, this tool will be effectively integrated into the workflow of daily clinical practice.

Conclusion

According to the methodology used in the present study, the Google Inception v3 model provided a high classification accuracy for tomographic images of OKCs and AMs. A higher rate of classification errors was observed for AM images compared to OKCs. Since this is the first study that used an automatic classifier to differentiate similar odontogenic tumors based on MDCT images, further investigations are necessary to elucidate the true accuracy of CNNs for the diagnosis of these tumors and to identify the best approach to introducing artificial intelligence in the daily practice of radiologists. It is strongly suggested that studies including other types of lesions histopathological lesions be performed. An analysis of other parametrizations and other types of CNNs could be made. The new approach to generate data augmentation, based on GANs (Generative Adversarial Networks), also is an interesting future research that can improve the learning process.