Application of neural networks to image recognition of wheat rust diseases

Poster (download) Mikhail Genaev11Mikhail Genaev ICG SB RAS, Kurchatov Genomic Center, Novosibirsk, Russia NSU, Novosibirsk, Russia, mag@bionet.nsc.ru Rust diseases of cereals are caused by pathogenic fungi and can significantly reduce plant productivity. Many cultures are subject to them. The disease is difficult to control on a large scale, so one of the most relevant approaches is crop monitoring, which helps to identify the disease at an early stage and make efforts to prevent its spread. One of the most effective methods of control is the identification of the disease from digital images that obtained by a smartphone camera. In this paper, we present a deep learning algorithm that uses a digital image of wheat plants to determine whether they are affected by a disease and, if so, what type: leaf rust or stem rust. The algorithm based on the convolution neural network of the densenet architecture. The resulting model demonstrates high accuracy of classification: the measure of accuracy F1 on the validation sample is 0.9, the AUC averaged over 3 classes is 0.98.

Read More

Named entity recognition in medical texts in Russian using deep learning models

Igor Viktorovich Moskalev1, Luybov\’ Anatol\’evna Khvorova21ASU, Barnaul, Russia, moskalev.igor.v@gmail.com2ASU, Barnaul, Russia, khvorovala@gmail.com The application of contextual and domain-specific pre-trained word embeddings for recognition of medical concepts in free-text clinical notes in Russian is considered. As it is known, a large amount of medical data is stored in electronic form, a significant part – in an unstructured form (medical history, extracts, description of the results of various tests). This data contains a large amount of useful information for the diagnosis of diseases. The results of the experiments showed the effectiveness of applying contextual language models which pre-trained on medical texts in the task biomedical named entity recognition.

Read More

The Siberian multimodal brain tumor image segmentation dataset (SBT)

Poster (download) Sergey Golushko1, Mikhail Amelin2, Bair Tuchinov3, Evgeniya Amelina4, Nikolay Tolstokulakov5, Evgeniy Pavlovskiy6, Vladimir Groza71Novosibirsk State University, s.k.golushko@gmail.com2FSBI \”Federal Neurosurgical Center\”, amelin81@gmail.com3Novosibirsk State University, bairts@gmail.com4Novosibirsk State University, amelina.evgenia@gmail.com5Novosibirsk State University, n.tolstokulakov@g.nsu.ru6Novosibirsk State University, pavlovskiy@post.nsu.ru7Median Technologies, vladimir.groza@gmail.com Automatic brain tumor segmentation from CT or MRI scans is one of the crucial problems among other directions and domains where daily clinical workflow requires to put a lot of efforts while studying patients with various pathologies.In this paper, we report the results of the research project \”Brain Tumor Segmentation\” organized in conjunction with the Federal Neurosurgical Center. Several state-of-the-art tumor segmentation algorithms were applied to a set of 100 MRI scans of meningioma, neurinoma and glioma patients – manually annotated by up to three raters – and to 100 comparable scans obtained using the automated tumor multi-region segmentation. Quantitative evaluations revealed a considerable agreement between the human raters in segmenting various tumor sub-regions (Dice scores in the range 85-90\\%). We found that different algorithms worked best for different sub-regions, but no single algorithm ranked in the top for all subregions simultaneously.

Read More