Neuroengineering-project

Vocal Tract Segmentation with U-net based framework from MRI images with overimposed Gaussian noise

Project from the course Neuroengineering @ Politecnico di Milano

• Mattia Cazzolla (@MattiaCazzolla) [email protected]

• Alix de Langlais (@Adelanglais) [email protected]

• Paolo Marzolo (@pollomarzo) [email protected]

• Olmo Notarianni ([email protected])

• Sara Rescalli ([email protected])

Final grade: 32/30

Dataset

The dataset provided was generated using the frames of Dynamic Supine MRI (dsMRI) videos recorded for different patients under specific speech protocols.
All the images had additive Gaussian noise overimposed.

The dataset contained a total of 820 images from 4 patients (respectively 280, 240, 150, 150).

Preprocessing

The preprocessing pipeline implemented aims at:

• Removing the Gaussian noise with a Total Variation Denoising technique (link)
• Enhancing the high frequency component

Model

The U-net architecture implemented consists of a variation from the IMU-NET described in this paper

Evaluation

The images were split into different datasets as follows:

• Patient 1 and 2 $\rightarrow$ Training Set
• Patient 3 $\rightarrow$ Validation Set
• Patient 4 $\rightarrow$ Test Set

The results on the test set are reported in the following table

Class	DICE (mean $\pm$ std)
Background	0.991 $\pm$ 0.001
Upper Lip	0.901 $\pm$ 0.033
Lower Lip	0.898 $\pm$ 0.018
Hard Palate	0.819 $\pm$ 0.045
Soft Palate	0.797 $\pm$ 0.059
Tongue	0.931 $\pm$ 0.012
Head	0.968 $\pm$ 0.007

The progress in learning can be observed by the segmentation at each epoch of the training

Video

The project required us to produce a 3 minutes video explaining our approach.

Untitled.mp4

Voiced by: @Adelanglais, @pollomarzo
Animated by: @Adelanglais, @MattiaCazzolla

Licence

This project is licensed under the MIT License.