# Deep learning-based approach for differential diagnosis of odontogenic cysts from histopathological images

## TL;DR

This paper introduces a deep learning system to help diagnose different types of tooth-related cysts using microscopic images, aiming to improve accuracy and speed up diagnosis.

## Contribution

A novel deep learning-based approach for differential diagnosis of odontogenic cysts using histopathological images is proposed.

## Key findings

- VGG16 and VGG19 models achieved the highest diagnostic accuracy (0.89) for odontogenic cyst classification.
- Xception and Inception V3 models showed slower convergence during training.
- Deep learning models can serve as effective decision support tools for diagnosing odontogenic cysts.

## Abstract

This study aims to provide Deep Learning (DL) based artificial intelligence (AI) methods using histopathology images to diagnose different types of odontogenic cysts (OCs) differentially.

Within the scope of the proposed study, hematoxylin and eosin (H&E) stained images of 3 different cyst groups were used. The dataset consists of histopathology images of 87 Dentigerous cysts (DC), 198 radicular cyst (RC), and 63 odontogenic keratocyst (OKC). Each image was zoomed with 3 different zoom levels and resized to 224x224 as preprocessing. In addition to the classical CNN method, Inception V3, VGG16, VGG19, and Xception architectures were used. The data set was split into training, validation, and test groups to avoid retesting.

The average accuracy, precision, sensitivity (recall), and F1-Score values obtained for CNN were 0.77, 0.80, 0.77, 0.75, and for VGG16 were 0.89, 0.90, 0.89, 0.89. For VGG19, these values were determined as 0.89, 0.90, 0.89, and 0.88, for Xception, these values were determined as 0.62, 0.52, 0.62,and 0.52 and for Inception, these values were determined as 0.62, 0.62, 0.62 and 0.56.

It was observed that VGG16 and VGG19 models showed superior performance on the data set in question, while Xception and Inception V3 models converged slower, meaning the training process progressed slower. Results showed that deep neural networks can be efficiently used in detecting OCs. AI-based OC detection may be a decision support tool that reduces interprofessional variability, expedites the diagnostic process, and lessens clinician workload.