posted on 2022-08-01, 00:00authored bySilvia Leccabue
Nowadays, research provides growing evidence about the importance of having tissue mechanical property values available as a parameter for diagnosis because tissue mechanical properties are sensitive to pathological changes. This behavior is the basis for using elastography as a diagnostic tool. The approach used to measure wave propagation in this work is a dynamic elastography technique called Scanning Laser Doppler Vibrometry (SLDV). Since several studies demonstrated the ability of Convolutional Neural Networks (CNNs) to automatically and robustly assess mechanical properties from elastography images overcoming the limitations of the traditional inversion techniques, this thesis aims to estimate the stiffness from silicon-based materials SLDV images through an implemented CNN. During the analysis, the use of soft tissue-mimicking materials as the EcoflexTM allowed us to perform the evaluations on some materials that realistically simulate the properties of soft tissues, exhibiting similar deformation responses and stiffness values. The classification of the shear modulus of the materials was performed on two separate tasks, a binary classification and a more complex five classes classification.
The proposed CNN architecture was pre-trained using synthetic wave data generated using a computational model and afterwards the model was fine-tuned with the physical data.
During the two experiments using physical data, the binary classification achieves an accuracy= 84.4%, and the multi-class classification reports an accuracy of 76,6%. CNN architecture is able to distinguish all the different materials, showing a significant rate of well-predicted data concerning both the simulation of healthy and pathological samples (binary classification) and also more intermediate stages of a specific pathology (five-classes classification), where the discriminant features are the changes in the wave propagation speed and the correlated shear modulus.
Although not yet allowing a clinical application for the estimation of the stiffness of organs and soft tissues, these results constitute a step forward towards the implementation of an
automatic and reliable method for assessing mechanical properties from elastography images.