Automated Classification of Coloboma Subtypes Using InceptionV3 Algorithm on Optical Coherence Tomography Images

Abstract

Introduction: Congenital coloboma is an eye abnormality in which genetically affected newborns have missing elements of the eye structure. That is why it is so important to correctly classify coloboma and then differentiate between the specific subtypes. Current diagnostic methods involve much dependence on the judgment of ophthalmologist andfundamentally rely on manual interpretation of OCT images which is cumbersome and unstandardized.

Objectives: The objective of this study is to create an automated classification system that will satisfactorily classify six forms of coloboma using OCT images with the help of the pre-trained InceptionV3 convolutional neural network. The objective in developing this tool is to have an optimally stable instrument that shall help clinicians in identifying and diagnosing various types of coloboma as quickly as possible.

Methods: Such a dataset of OCT images of the coloboma conditions included various subtypes of coloboma. These images were rescaled and normalized, and pre-process by enlarging or shrinking it to the same size as the target input images. The used model in this study is the InceptionV3 model which was pre-trained on ImageNet and subsequently fine-tuned on this dataset. To reduce overfitting, rotation, scaling, and flip were performed as approach to data augmentation. Independent data validation was done to measure the accuracy of the model and using sensitivity, specificity, and the AUC-ROC curve.

Results: The InceptionV3 model trained with a narrowed focus was accurate in recognising the various coloboma subtypes and reduced the diagnostic time as well as precision more effectively. This automated classification system can be useful to clinical decision making and patient care and should be implemented as a part of different screening systems.

Conclusions: The evaluated proposed automated colos subtype classification system based on optical coherence tomography imaging and the InceptionV3 model can contribute to the improvement of clinical results due to more accurate and fast subtype classification. Further work will include the collection of more data, the inclusion of more imaging data and the enhancement of efficiency of the model and its results.