Abstract
Purpose :
Several supervised machine learning models have been proposed to assist in keratoconus (KCN) detection, however, they require numerous well annotated data. Herein, we propose a new unsupervised hybrid artificial intelligence model to detect KCN.
Methods :
We have developed an unsupervised model based on the k-means algorithm and adapted flower pollination algorithm (FPA). The proposed model was evaluated using two independent datasets: Pentacam (Oculus Inc., Germany) corneal imaging data collected from 5430 eyes from Department of Ophthalmology and Visual Sciences, Paulista Medical School, Federal University of Sao Paulo, Sao Paulo, Brazil; and CASIA (Tomey, Japan) corneal imaging data of 1531 eyes from Department of Ophthalmology, Jichi Medical University, Tochigi in Japan. Three different clustering scenarios were evaluated: 2 classes (normal vs KCN), 3 classes (normal vs KCN low grade (1-2 stage) vs KCN high grade (3-4 stage) and 5 classes (normal, stage KCN 1, KCN 2, KCN 3 and KCN 4). We compared the proposed method with three other standard unsupervised algorithms including K-means alone, K-medoids, and Spectral cluster (figure 1), following 25 runs. Accuracy (Ac) metrics analyses also included Precision (Pr), Recall (R), F-Score (F), and Purity (Pu).
Results :
FPA-K-means outperformed the other algorithms in 2-, 3- and 5-class scenarios, except for Recall and Purity for the 5-class scenario. In the 2-class test, FPA-K-means achieved Ac=96.03, Pr=96.29, R=96.06, F=96.17, and Pu=96.03. When considering a 3-class scenario, FPA-K-means reached Ac=71.02, Pr=53.53, R=75.37, F=64.64, and Pu=80.85. While for the 5-class test, FPA-K-means attained Ac= 75.20, Pr= 35.01, R= 47.70, F= 48.97, and Pu= 53.64, whereas Spectral clustering method achieved R= 55.33 and Pu= 55.12
Conclusions :
The proposed model consists of a new unsupervised algorithm that combines two computational learning techniques for detecting KCN from corneal image data. As it is unsupervised, the model is not affected by expert pre-labelling bias. Our study demonstrates that the proposed model can be used in the clinical and scientific setting for detection of KCN. Improvements to this model may be applied in clinics for diagnosing the corneal status of KCN patients.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.