guest ::
| Home > Publications database > Modifying an existing Convolutional Neural Network to predict Total Intracranial Volume using T1w images |
| Home > Publications database > Modifying an existing Convolutional Neural Network to predict Total Intracranial Volume using T1w images |
| Coursework | FZJ-2024-01138 |
; ; ;
2023
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2024-01138
Abstract: Neuroimaging has become an essential part in diagnosing and treating neurological diseases.Statistics such as sex, age, and dexterity can be asked for by medical staff during consultation. Biological markers or medical information such as total intracranial volume (colloq.: brain volume; TIV) or absolute and relative amount of grey/white matter have to be computed by usually timeconsuming calculations.It is of both patient’s and medical staff’s interest to have results as soon as possible after collecting data from magnetic resonsance imaging (MRI), a type of imaging procedure.Machine Learning provides statistical tools to omit classical time-consuming processing and make predictions about those biological markers based on previously analyzed brain images. This allows faster estimates and enables timely discussion and diagnosis with and of the patient after examination.Especially in emergency situations this can provide doctors with required information and allows them to start treatment while the patient is still on-site, if necessary.There have been previous efforts on building a Convolutional Neural Network (CNN) that is able to predict brainage on a given T1-weighted image. T1w images are a type of neuroimaging data.This thesis will describe modifications done to the given CNN ’Simple Fully Convolutional Network’ (SFCN), in order to predict TIV using the same T1w image of the brain as input.It will furhter explain various Machine Learning concepts and the SFCN’s implementation as well as the training and evaluation functions used as the foundation for this appliance.This modification is done by linear scaling the given TIV values used for back- propagation into the space the SFCN has previously been trained and applied on, in this case the age range of [42, 82]. This has been implemented into the code of SFCN.After modifying, the model has been trained for 10 epochs using 1370 T1w brain images provided by the Amsterdam Open MRI Collection (AOMIC)Using the Kullback-Leibler divergence as a measure of loss, the model shows a training loss of 2.46 after 1 epoch and 1.98 after 10 epochs.Its validation loss initially starts at 3.4 while ending at 3.45 after 10 epochs. The validation loss is minimal after 2 epochs with a KL-divergence of 2.7.
Keyword(s): Medicine (2nd)
|
The record appears in these collections: |
PDF