AIMING: Unbiased and explainable artificial intelligence for medical imaging

Nowadays, the relevance that Artificial Intelligence (AI) technologies will have in our society has been clearly foreseen by policymakers at all scales. When listing the envisaged domains that will benefit from the use of AI-enabled technologies, healthcare is always in the first positions. In this context, medical imaging is one of the most promising applications of artificial intelligence in healthcare.

Many robust and advanced Machine Learning / Deep Learning (ML/DL) algorithms are available for image analysis, but their validation and implementation in the medical domain is still limited because of several challenges. In the AIMING project, we will address two of these main technological challenges that prevent the application of ML/DL advances in medical imaging: 

Biases in the models due to training data: ML/DL models in medical imaging face several challenges due to potential biases introduced by

(i) data availability, as no data is available or only small datasets exist; (ii) lack of annotated or labelled data, which must be annotated by experts; (iii) dataset shift, as models may not generalize to different domains; and (iv) evolving datasets in deployment, as data may vary due to changes in acquisition protocols.

Explainability of the results: Many ML/DL models do not provide an interpretation of their outcome. This insight may help to better understand a disease, to unveil new biomarkers or provide clues for a finer stratification of patients, becoming tools towards a more personalized medicine. Another aspect is the assessment of the results uncertainty. It can be used to understand which specific samples

are identified as being difficult and may require closer examination. 

The basic research developed when facing the previous challenges will be put in practice in the context of three different medical domains. In each one of these domains, the AIMING project will tackle a socially relevant disease.

In the context of histological imaging, we will work with different kinds of inmunohistochemical stains for breast cancer diagnosis and type of chemotherapy selection. The objectives are (i) to improve results by introducing adaptive learning in the annotation system, (ii) to generalize the results to other acquisition systems and to different tumors, and (iii) to determine the tumor areas in the H&E stained tissue

section. This work will be carried out in collaboration with seven major hospitals of the Institut Català de la Salut (ICS).

In the neuroimaging domain, we will study the Alzheimers disease (AD). There are two main goals: (i) developing a pre-screening tool for detecting, at early stages, subjects at risk of Alzheimers disease by combining MRI-derived features and additional features like demographics (age, sex, education, cognitive tests) and genetic (apoe4 status) and, (ii) predicting biological age to implement transfer learning techniques between age and pre-AD prediction. This research will be developed in collaboration with the Barcelona Brain Research Center (BBRC).

In the dermoscopic imaging area, we will focus on melanoma detection. We will work on two main aspects: (i) to develop more performing models for melanoma detection, especially in unconstrained settings and using multimodal and contextual information and, (ii) to build a survival prediction system for melanoma patients towards personalized medicine. This research will be conducted in collaboration with the Melanoma Unit at IDIBAPS.

Acknowledgements for publications:

This work has been supported by the Spanish Research Agency (AEI) under project PID2020-116907RB-I00 of the call MCIN/ AEI /10.13039/501100011033