Domain-specific loss design for unsupervised physical training: A new approach to modeling medical ML solutions
Authors:
Hendrik Burwinkel,
Holger Matz,
Stefan Saur,
Christoph Hauger,
Ayse Mine Evren,
Nino Hirnschall,
Oliver Findl,
Nassir Navab,
Seyed-Ahmad Ahmadi
Abstract:
Today, cataract surgery is the most frequently performed ophthalmic surgery in the world. The cataract, a develo** opacity of the human eye lens, constitutes the world's most frequent cause for blindness. During surgery, the lens is removed and replaced by an artificial intraocular lens (IOL). To prevent patients from needing strong visual aids after surgery, a precise prediction of the optical…
▽ More
Today, cataract surgery is the most frequently performed ophthalmic surgery in the world. The cataract, a develo** opacity of the human eye lens, constitutes the world's most frequent cause for blindness. During surgery, the lens is removed and replaced by an artificial intraocular lens (IOL). To prevent patients from needing strong visual aids after surgery, a precise prediction of the optical properties of the inserted IOL is crucial. There has been lots of activity towards develo** methods to predict these properties from biometric eye data obtained by OCT devices, recently also by employing machine learning. They consider either only biometric data or physical models, but rarely both, and often neglect the IOL geometry. In this work, we propose OpticNet, a novel optical refraction network, loss function, and training scheme which is unsupervised, domain-specific, and physically motivated. We derive a precise light propagation eye model using single-ray raytracing and formulate a differentiable loss function that back-propagates physical gradients into the network. Further, we propose a new transfer learning procedure, which allows unsupervised training on the physical model and fine-tuning of the network on a cohort of real IOL patient cases. We show that our network is not only superior to systems trained with standard procedures but also that our method outperforms the current state of the art in IOL calculation when compared on two biometric data sets.
△ Less
Submitted 9 May, 2020;
originally announced May 2020.
