Differentiation between anterior segment images of femtosecond laser-assisted and manual cataract surgery utilising artificial intelligence

Session Details

Session Title: Femtolaser-Assisted Cataract Surgery (FLACS)

Session Date/Time: Monday 16/09/2019 | 08:30-10:30

Paper Time: 10:10

Venue: Free Paper Forum: Podium 4

First Author: : K.Yamazaki JAPAN

Co Author(s): :    R. Kimoto   Y. Shibata   J. Yoneyama   T. Hayashi   T. Mimura   H. Takahashi              

Abstract Details

Purpose:

To use artificial intelligence (AI) to objectively analyze the difference in post-operative architecture between femtosecond laser-assisted cataract surgery (FLACS) and manual cataract surgery using anterior ocular segment images.

Setting:

Omiya Nanasato Eye Institute, Saitama, Japan

Methods:

Capsulotomy, lens fragmentation, and corneal incisions were performed with FLACS. This investigation involved 254 patients and 2,005 images, using an anterior ocular segment slit lamp from postoperative Day 1 to 6 months. FLACS with multifocal, FLACS with monofocal, Manual with multifocal, and Manual with monofocal IOL surgeries were performed for 70, 44, 70, and 70 patients, respectively. The best deep learning architecture was selected from various patterns of the GoogLeNet and the ResNet deep learning architectures. Ten-fold cross-validation was used to evaluate the models. Images from each patient were grouped in separate batches.

Results:

If the architectures were only random, the accuracy would be 25%. However, the accuracy of FLACS with multifocal, FLACS with monofocal, Manual with multifocal, and Manual with monofocal was 54, 39, 63, and 47%, respectively. The accuracy rate of FLACS was 66%, being lower than the accuracy rate of multifocal IOL at 69% (P < 0.0001). Although the odds ratio between FLACS and multifocal IOL was 2.2, manual with multifocal was misestimated as FLACS with multifocal IOL in only 12% of responses. IOL type alone did not allow the architecture to associate surgery with FLACS.

Conclusions:

AI statistically differentiated anterior segment images of FLACS from manual surgery. Our results objectively show that the accuracy of capsulotomy and wound architecture is more precise with FLACS than with manual surgery.

Financial Disclosure:

None