Abstract
Purpose :
Using in vivo confocal microscopy (IVCM) is an efficient way to diagnose ocular surface squamous neoplasia (OSSN). Convolutional neural networks (CNNs) can work with high accuracy for image classification problems, but always require a large number of training samples. Here, we provide a method to achieve high accuracy with a small dataset and a model that can predict OSSN with high confidence.
Methods :
Our dataset was composed of 2,774 corneal and conjunctival IVCM images of five classes: OSSN, normal cornea, melanoma, pterygium and keratitis. 745 images had OSSN lesions, 2,029 images had non-OSSN lesions. Healthy and unhealthy patterns are clearly distinguishable, but in pathological cases the different symptoms may have similar image features that must be annotated by experts. Image annotation was performed manually by identifying the abnormal regions, according to the following criteria: 1) Starry-sky pattern (hyperreflective nucleus), 2) hyperkeratosis, 3) mitotic figures, 4) enlarged, irregular epithelial cells. Collecting labeled regions instead of using the entire image as information can significantly extend the size of our training set but the patterns are often patient-specific. In order to have a patient-independent and diverse training set, we proposed to elaborate a cell-level classification that can imply the diagnosis.
Results :
In the binary classification the highest accuracy was 95%, precision 93%, recall 100% and 95% F1 score. The model achieved the highest accuracy of 91% in the four-class labeled classification, with a precision of 86%, recall of 79% and F1 score of 82%. To implement the cell-level classification, preliminary annotation was made using cell segmentation tools.
Conclusions :
We established a technique and a deep learning model for detecting alterations on IVCM images characteristic to OSSN having a small dataset. Our network demonstrated a high accuracy in binary classification and pattern recognition but the results can be further improved if the classification is extended to cellular level. To the best of our knowledge, this study is the first to develop and test an artificial intelligence model to detect OSSN using IVCM images.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.