Supervisors info:
Ευστάθιος Στυλιάρης, Καθηγητής, Τμήμα Φυσικής, ΕΚΠΑ
Summary:
This Bachelor’s thesis aims to approach medical imaging through the use of
artificial neural networks. Medical physics and especially medical imaging plays a
crucial role in modern medicine, reinforcing diagnostic and therapeutic process through
advanced technologies and are therefore, a useful tool in the early diagnosis and
monitoring of the progress of neurodegenerative brain diseases such as Alzheimer's
disease.
Positron emission tomography (PET) and single photon emission tomography
(SPECT) are the main applications of nuclear medicine that use radiotracers to evaluate
biological processes in living organisms and are both used in the diagnosis and
assessment of the prognosis of Alzheimer’s disease. One tomoscintigraphic technique
based on SPECT imaging is brain scintigraphy (DaTSCAN), which detects dopamine
transporter activity in the brain and is used to distinguish patients with dementia with
Lewy bodies from patients with Alzheimer's disease and to distinguish patients with
Parkinson’s disease from those with Alzheimer’s disease. A normal DaTSCAN image
includes equal absorption of the radiotracer by the two brain lobes which are displayed
in an elliptical shape. In the case of patients with Alzheimer’s disease, the DaTSCAN
image shows mild alterations as the uptake of the radiopharmaceutical is not greatly
affected by the disease, in contrast to other neurodegenerative diseases in which the
alterations are more intense. In any case, the diagnosis of the respective disease is
made by the registered neurologist.
As an initial approach to neural networks, simple neural networks are being
presented that generate the required sample of images with dimensions of 128 x128
and then classify them according to specific features that are the the task of each
problem. The samples of images include simple geometric shapes such as squares and
circles and display some variations in their position, size, multiplicity and overlap.
Based on these characteristics, the image classification into categories is requested
according to the predefined task. Through the mentioned simple neural networks, the
process of training is being studied and it is found that by setting the appropriate
conditions in training and the appropriate parameters in the network architecture,
neural networks are able to generalize to unknown data and correctly categorize the
images, with accuracy rates ranging from 82.5% up to 100%.
In the scope of approaching medical imaging through the use of artificial
intelligence, a convolutional neural network is being developed that accepts as input
data phantoms of tomoscintigraphic brain images and has as output the quantified
absorption of the radiopharmaceutical by the two brain lobes. Another convolutional
neural network is also being developed with the same input data and has as output the
absorbed intensities od each brain lobe seperately as well as the absorbed intensity of
the background. For the construction of the phantoms, the Simulix3x program is used to
generate the required sample images with dimensions of 128 x128, containing the
head ellipsoid and the two lobe ellipsoids with a stochastic shift in the position and
intensity of the absorbed radiation from both the lobes and the background. A total of
3000 phantoms were generated with this software, of which 80% were used for model
training, while the remaining 20% were used to monitor the performance and
convergence of the neural network. The two convolutional neural networks mentioned above present standard
deviations of σ1 = 0,014 for the quantified absorption of the radiopharmaceutical by the
two brain lobes and σ2= 0,078 for the prediction of the absorbed intensity of each brain
lobe. Regarding to the prediction for the absorbed intensity of the background, the
standard deviation is equal to σ3 = 0,025. These standard deviations as well as the
convergence of the convolutional neural networks indicate that these artificial neural
networks can be applied to real medical data.
Finally, the neural networks are evaluated on two real DaTSCAN medical images and
the percentage deviation of the predicted values compared to the values obtained from
the tomoscintigraphic images is calculated, which is on the order of 4.6% for the
absorbed intensities of the two cerebral lobes. In conclusion, considering that this
deviation is within the acceptable range for medical applications, it is possible to use
the convolutional neural network as an adjunct in diagnosis and in decision making by
the treating medical specialist.
