Deep Convolutional Autoencoders for reconstructing magnetic resonance images of the healthy brain

Deep Autoencoder for brain MRI

---

Master's Thesis. Master's in Data Science at Universitat Oberta de Catalunya.

Author

• Adrián Arnaiz Rodríguez - aarnaizr@uoc.edu

Tutor:

• Dr. Baris Kanber - bkanber@uoc.edu

---

• Report (82 pages)
• Slides (20 slides)
• Papers with code

---

Convolutional Autoencoder Architectures used

Experiments:

• With Data Augmentation:

  • 2 experiments (MSE and DSSIM Loss) for each of the following architectures:
    • Shallow residual autoencoder (full-pre)
    • Shallow residual autoencoder (full-pre) + L2 reg.
    • Skip connection autoencoder
    • Skip connection autoencoder + L2 reg.
    • Myronenko Autoencoder
    • RESIDUAL-UNET (proposed new improved architecture)

• Without Data Augmentation:

  • MSE Loss
    • Shallow residual autoencoder (original)
    • Shallow residual autoencoder (full-pre)
    • Shallow residual autoencoder (full-pre) + L2 reg.
    • Skip connection autoencoder
    • Myronenko Autoencoder
    • Myronenko Autoencoder + L2 reg.

---

Results

Quantitative Results

Model	loss	L2	Val loss	MSE	DSSIM	PSNR
Residual U-NET	MSE	No	3.58e-05	3.44e-05	2.95e-03	44.9
Shallow RES full-pre	MSE	No	1.55e-04	1.51e-04	6.75e-03	38.6
Skip connection CAE	MSE	Yes	2.69e-04	2.25e-04	1.65e-02	36.8
Skip connection CAE	MSE	No	3.10e-04	2.99e-04	9.36e-03	35.7
Myronenko CAE	MSE	No	3.38e-04	3.27e-04	1.57e-02	35.1
Shallow RES full-pre	MSE	Yes	3.72e-04	3.24e-04	1.14e-02	35.2
Residual U-NET	DSSIM	No	1.50e-03	7.49e-05	1.44e-03	41.8
Shallow RES full-pre	DSSIM	Yes	4.42e-03	2.34e-04	3.70e-03	36.7
Shallow RES full-pre	DSSIM	No	4.19e-03	2.88e-04	4.14e-03	35.9
Myronenko CAE	DSSIM	No	4.39e-03	6.69e-04	4.31e-03	32.1
Skip connection CAE	DSSIM	Yes	4.82e-03	4.08e-04	4.38e-03	34.2
Skip connection CAE	DSSIM	No	4.90e-03	4.57e-04	4.71e-03	33.7

Quanlitative Results

Char of test metrics and dependent sample t-test significances