Bridging physics and deep learning: an autoencoder-based approach for non-invasive estimation of prostate tissue composition

Batuhan Gundogdu; Aritrick Chatterjee; Serene Bose; Gregory S. Karczmar; Aytekin Oto

doi:10.1117/12.3006626

1 April 2024 Bridging physics and deep learning: an autoencoder-based approach for non-invasive estimation of prostate tissue composition

Batuhan Gundogdu, Aritrick Chatterjee, Serene Bose, Gregory S. Karczmar, Aytekin Oto

Author Affiliations +

Proceedings Volume 12925, Medical Imaging 2024: Physics of Medical Imaging; 129254N (2024) https://doi.org/10.1117/12.3006626
Event: SPIE Medical Imaging, 2024, San Diego, California, United States

Conference Poster

Abstract

Physics model-based approaches using Magnetic Resonance Imaging (MRI) have been developed to provide a non-invasive tool to estimate tissue composition by fitting the notoriously ill-conditioned sum-of-decaying-exponentials functions. However, even modest levels of noise can negatively impact the accuracy of such models. This study introduces a new paradigm that integrates the strengths of physics model-based and deep learning-based methods, while overcoming their respective weaknesses. We reformulated the MRI-based tissue composition estimation using a physics-informed autoencoder (PIA), transforming the problem from a least-squares fit into a robust deep learning framework. The PIA model consists of a trainable multi-head neural network encoder and a fixed, physics-informed decoder. Without a need for further fine-tuning for each new patient, the encoder’s latent activations are utilized to derive prostate tissue composition estimates, such as volume fractions for each compartment along with their diffusivity and T2. PIA’s estimates of tissue parameters are juxtaposed with the least-squares solution through both Monte Carlo simulations and the quantitative histology of in-vivo prostate scans. In-vivo evaluations, performed on 19 pathology-proven prostate cancer patients, have been validated with quantitative histology-based true tissue compositions. Both Monte Carlo and in-vivo evaluations reveal that PIA significantly outperforms the conventional least-squares solution on Pearson correlation and mean absolute error metrics. Notably, PIA was able to predict the diffusivity and T2 of each tissue compartment more accurately, whereas the least-squares method tends to pin these values to the extremities.

(2024) Published by SPIE. Downloading of the abstract is permitted for personal use only.

Citation Download Citation

Batuhan Gundogdu, Aritrick Chatterjee, Serene Bose, Gregory S. Karczmar, and Aytekin Oto "Bridging physics and deep learning: an autoencoder-based approach for non-invasive estimation of prostate tissue composition", Proc. SPIE 12925, Medical Imaging 2024: Physics of Medical Imaging, 129254N (1 April 2024); https://doi.org/10.1117/12.3006626

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available