(c) Tonpor Kasa
Researchers from CEA-Joliot (NeuroSpin), in collaboration with the Gustave Roussy Institute, the Necker Hospital and the Curie Institute (Orsay), are proposing an original method for analyzing MRI images of rare brain tumors, by combining detection automatique (Automatic control is part of the engineering sciences. This discipline deals with the…) of objects and segmentation (In general, the word segmentation designates the action of segmenting, the fact of segmenting oneself…) by learning (Learning is the acquisition of know-how, i.e. the process…) deep on common tumors.
To study brain tumors from MRI images, oncologists must precisely delineate the contours of the lesions (delineation) or “segment” them, i.e. group the pixels of the image into different groups, according to predefined criteria.
There are currently several architectural (Architectures is a documentary series proposed by Frédéric Campain and Richard Copans,…) deep learning segmentation for brain tumors. These models are only effective for the types of tumor (The term tumor (from the Latin tumere, to swell) designates, in medicine, an increase in…) on which they were trained. They are therefore better for common tumors, such as glioblastoma multiforme (Glioblastoma multiforme (GBM), also known as grade 4 astrocytoma, is the…)than for rare tumours, such as glioma (Gliomas or glial tumors are all brain tumors arising from the supporting tissue or…) infiltrating trunk (A trunk can be:) brain (pediatric cancer).
However, there are some visual similarities between common and rare tumors that allow the problem to be approached in two steps: detection and then pixel classification.
This is the approach adopted by NeuroSpin researchers and their partners. They offer two delineation methods based on:
– object detection with an automatic object detection algorithm, known for its high precision and speed (You Only Look Once);
– the segmentation of tumors with a network (A computer network is a set of equipment linked together to exchange information…) of convolutional neurons, developed for the processing of biomedical images (U-Net).
For each step, neural networks trained on common lesions were used on rare lesions, without adjusting additional parameters. This strategy (Strategy – from the Greek stratos which means “army” and ageîn which means…) allowed to obtain better results when the tumor differs from that of the training and robust delineations were obtained on the infiltrating glioma of the brainstem.
By tackling the issue of rare tumours, for which no database (In computing, a database (Abr.: “BD” or…) cannot be built to train a deep neural segmentation network, the researchers show that by combining “simple” object detection and tumor segmentation, good results can be obtained, without retraining or adaptation of the model.
Object Detection Improves Tumor Segmentation in MR Images of Rare Brain Tumours, Cancers
Did you like this article? Do you wish to support us ? Share it on social networks with your friends and/or comment on it, this will encourage us to publish more similar topics!