Portal de Periódicos da CAPES

Influence of Cross Histology Transfer Learning on the Accuracy of Medical Diagnostics Systems

2023; Springer International Publishing; Linguagem: Inglês

10.1007/978-3-031-27499-2_86

2367-3370

Alexander Mongolin, Sergey G. Khomeriki, Nikolay Karnaukhov, Konstantin Abramov, Roman Vorobev, Yuri Gorbachev, Anastasia Zabruntseva, Alexey Kornaev,

Digital Imaging for Blood Diseases

Transfer learning is a basic method in computer vision that helps extract features from images. In this research, we created a private data set for stomach cancer recognition and tested the various transfer learning techniques in an application to solution the binary classification problem. The obtained results show that it is not always necessary to use pretrained weights of similar nature (histology) or ImageNet-based models. Given enough data, randomly initialized weights can show more promising results. Meanwhile, the goal of the network pretrained with data set of different type of cancer is relatively high precision of the network in comparison with the network pretrained with ImageNet or the randomly initialized network.