Keratoconus severity identification using unsupervised machine learning

• DOI: 10.1371/journal.pone.0205998
• Publication Date: 2018-11-06T18:30:29Z
• Authors: Siamak Yousefi; Ebrahim Yousefi; Hidenori Takahashi; Takahiko Hayashi; Hironobu Tampo; Satoru Inoda; Yusuke Arai; Penny Asbell
• Keywords: Male; Principal Component Analysis; Science; Q; R; Keratoconus; Severity of Illness Index; Cornea; 03 medical and health sciences; 0302 clinical medicine; Medicine; Humans; Female; Algorithms; Research Article; Aged; Dilatation, Pathologic; Unsupervised Machine Learning

We developed an unsupervised machine learning algorithm and applied it to big corneal parameters identify monitor keratoconus stages. A dataset of swept source optical coherence tomography (OCT) images 12,242 eyes acquired from SS-1000 CASIA OCT Imaging Systems in multiple centers across Japan was assembled. total 3,156 with valid Ectasia Status Index (ESI) between zero 100% were selected for the downstream analysis. Four hundred twenty topography, elevation, pachymetry (excluding ESI Keratoconus indices) selected. The included three major steps. 1) Principal component analysis (PCA) used linearly reduce dimensionality input data 420 eight significant principal components. 2) Manifold further reducing components nonlinearly two eigen-parameters. 3) Finally, a density-based clustering eigen-parameters keratoconus. Visualization clusters 2-D space validate quality subjectively assess accuracy identified objectively. proposed method four clusters; I: cluster composed mostly normal (224 equal zero, 23 five 29, nine greater than 29), II: healthy forme fruste (1772 698 117 III: mild stage (184 74 6 zero), IV: advanced (80 had 29 1 eye 29). found that status severity can be well using algorithms along linear non-linear transformation. better visualize