RESEARCH

Numerous approaches to cell-type deconvolution have been proposed, yet very few take advantage of the emerging discipline of deep learning and most approaches are limited to input data regarding the expression profiles of the cell types in question.

We have developed DECODE, a deep learning method for the task that is data-driven and does not depend on input expression profiles.

DECODE builds on a deep unfolded non-negative matrix factorization technique. It is shown to outperform previous approaches on a range of synthetic and real data sets, producing abundance estimates that are closer to and better correlated with the real values.

Despite extensive research efforts, brain tumors remain a leading cause of cancer-related death, with only one-third of individuals surviving more than 5 years after diagnosis. Early detection constitutes a decisive factor in disease prognosis. While various medical imaging modalities can provide an anatomical view of the brain, they only capture morphological tissue changes that occur relatively late, when the tumor is already well-defined and mature.

In contrast, alterations in metabolic properties, such as increased glucose consumption, are manifested as early as cell reprogramming and constitute a unique tumor signature. As currently employed in-vivo metabolic imaging techniques require the use of ionizing radiation and have limited spatial resolution, there is an urgent need for an alternative accurate, and safe means for early detection and characterization of brain tumors.

Recently we discovered a new non-toxic sugar-based material, which is preferentially accumulated in tumor cells and can be detected using the chemical exchange saturation transfer magnetic resonance imaging (CEST-MRI) technique. However, the complex molecular brain environment generates a variety of confounding magnetic signals stemming from various metabolites, lipids, and peptides, which severely hinder the quantification of metabolic sugar consumption.

The main goal of this project is to develop a transformative radiation-free, and rapid AI-based metabolic MR technology for early brain tumor detection. We propose to adopt a previously unconsidered perspective and to represent the underlying physics of molecular MRI as a computational graph, enabling an automatic AI-based optimization of tumor metabolic imaging.

Our research is centered on Explainable AI (XAI), wherein we aim to establish mathematical foundations for elucidating model predictions. 

The overarching goal is to discern the limitations and possibilities associated with explanations generated by AI models. 

By focusing on the development of robust methods, our work contributes to the broader objective of enhancing transparency, accountability, and trustworthiness in AI systems. 

The specific applications of this research span diverse fields such as medicine, healthcare, law, and ethics. 

Example from Our Research:

One of the challenges in machine learning, particularly in medical applications, is accurately interpreting feature importance scores, especially when complex dependencies exist in the data. 

A key situation that can lead to unexpected outcomes is the presence of a collider variable, also known as an inverted fork.

For instance, consider the example depicted in the attached Figure. Smoking (which is unobserved) increases the likelihood of both Cancer and Gum chewing. Additionally, doctors recommend chewing gum to patients with Earache. When building predictive models for Cancer using features like Gum and Earache, but without knowledge of Smoking, misleading conclusions may arise.

• Scenario 1: A model using only Earache as a feature would assign no importance to Earache since it is independent of Cancer.
   
• Scenario 2: A model using only Gum would show Gum as an important predictor of Cancer due to their correlation, even though this association arises from the unobserved variable (Smoking).
   
• Scenario 3: In a model using both Earache and Gum, feature importance scores might misleadingly indicate that Earache is predictive of Cancer. This is because conditioning on Gum (the collider) creates a spurious association between Earache and Cancer.

These kinds of situations, known in causality literature, are not easily resolved without additional information about the underlying causal mechanisms. 

In medical contexts, such erroneous interpretations could lead to incorrect conclusions, such as associating Earache with cancer risk or wrongly attributing gum chewing as a preventive factor for cancer. 

Our research explores both the theoretical underpinnings of these phenomena and provides empirical results to help mitigate such risks when applying machine learning models in healthcare