Prediction of Pentacam image after corneal cross-linking by linear interpolation technique and U-NET based 2D regression model

FIRAT M., ÇINAR A., ÇANKAYA C., Firat I. T., TUNCER T.

COMPUTERS IN BIOLOGY AND MEDICINE, cilt.146, 2022 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 146
  • Basım Tarihi: 2022
  • Doi Numarası: 10.1016/j.compbiomed.2022.105541
  • Dergi Adı: COMPUTERS IN BIOLOGY AND MEDICINE
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Applied Science & Technology Source, BIOSIS, Biotechnology Research Abstracts, CINAHL, Compendex, Computer & Applied Sciences, EMBASE, INSPEC, Library, Information Science & Technology Abstracts (LISTA), MEDLINE
  • Anahtar Kelimeler: Corneal cross-linking, Linear interpolation, 2D regression, Deep learning, Keratoconus, ULTRAVIOLET, RIBOFLAVIN
  • İnönü Üniversitesi Adresli: Evet

Özet

Keratoconus is a common corneal disease that causes vision loss. In order to prevent the progression of the disease, the corneal cross-linking (CXL) treatment is applied. The follow-up of keratoconus after treatment is essential to predict the course of the disease and possible changes in the treatment. In this paper, a deep learningbased 2D regression method is proposed to predict the postoperative Pentacam map images of CXL-treated patients. New images are obtained by the linear interpolation augmentation method from the Pentacam images obtained before and after the CXL treatment. Augmented images and preoperative Pentacam images are given as input to U-Net-based 2D regression architecture. The output of the regression layer, the last layer of the U-Net architecture, provides a predicted Pentacam image of the later stage of the disease. The similarity of the predicted image in the final layer output to the Pentacam image in the postoperative period is evaluated by image similarity algorithms. As a result of the evaluation, the mean SSIM (The structural similarity index measure), PSNR (peak signal-to-noise ratio), and RMSE (root mean square error) similarity values are calculated as 0.8266, 65.85, and 0.134, respectively. These results show that our method successfully predicts the postoperative images of patients treated with CXL.