Naslov
Retinal OCT Image Segmentation: How Well do Algorithms Generalize or How Transferable are the Data?

Autori
Melinščak, Martina ; Lončarić, Sven

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), znanstveni

Izvornik
MIPRO 2020 43rd International Convention Proceedings / Skala, Karolj - Rijeka : Hrvatska udruga za informacijsku i komunikacijsku tehnologiju, elektroniku i mikroelektroniku - MIPRO, 2020, 108-112

Skup
43rd International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO 2020)

Mjesto i datum
Opatija, Hrvatska, 28.09.2020. - 02.10.2020

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
deep learning ; retinal OCT images ; image segmentation ;

Sažetak
The success of deep learning depends, among other things, on a large amount of labeled data. However, in medical applications, large labeled datasets are the exception, rather than the rule. Manual image labeling is time-consuming and is generally performed only with the purpose of developing algorithms, and not as a part of standard clinical practice. The goal of this study is twofold. Since there is always a trade-off between the ability to collect data and achieve the best possible performance, we wanted to explore how performance depends on the amount of data. For this purpose, a database of manually annotated OCT images was collected. Also, we wanted to see how much transfer learning can help. Retinal OCT images vary depending on the type of device, therefore developed methods should be as robust as possible. Transfer learning was performed so that the model was trained with similar OCT images and then fine-tuned with images from the collected database. It has been shown that transfer learning helps in terms of generalization and better prediction in case the source database is similar to the target database. We can also assume that further improvement can probably be achieved by adding images from another distribution (medical or nonmedical).

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Interdisciplinarne tehničke znanosti

POVEZANOST RADA