Towards an effective and explainable AI: studies in the biomedical domain

Abstract

Providing accurate diagnoses of diseases and maximizing the effectiveness of treatments requires, in general, complex analyses of many clinical, omics, and imaging data. Making a fruitful use of such data is not straightforward, as they need to be properly handled and processed in order to successfully perform medical diagnosis. This is why Artificial Intelligence (AI) is largely employed in the field. Indeed, in recent years, Machine Learning (ML), and in particular Deep Learning (DL), techniques emerged as powerful tools to perform specific disease detection and classification, thus providing significant support to clinical decisions. They gained a special attention in the scientific community, especially thanks to their ability in analyzing huge amounts of data, recognizing patterns, and discovering non-trivial functional relationships between input and output. However, such approaches suffer, in general, from the lack of proper means for interpreting the choices made by the learned models, especially in the case of DL ones. This work is based on both a theoretical and methodological study of AI techniques suitable for the biomedical domain; furthermore, we put a specific focus on the practical impact on the application and adaptation of such techniques to relevant domain. In this work, ML and DL approaches have been studied and proper methods have been developed to support (i) medical imaging diagnostic and computer-assisted surgery via detection, segmentation and classification of vessels and surgical tools in intra-operative images and videos (e.g., cine-angiography), and (ii) data-driven disease classification and prognosis prediction, through a combination of data reduction, data visualization and classification of high-dimensional clinical and omics data, to detect hidden structural properties useful to investigate the progression of the disease. In particular, we focus on defining a novel approach for automated assessment of pathological conditions, identifying latent relationships in different domains and supporting healthcare providers in finding the most appropriate preventive interventions and therapeutic strategies. Furthermore, we propose a study about the analysis of the internal processes performed by the artificial networks during classification tasks, with the aim to provide a AI-based model explainability. This manuscript is presented in four parts, each focusing on a special aspect of DL techniques and offering different examples of their application in the biomedical domain. In the first part we introduce clinical and omics data along with the popular processing methods to improve the analyses; we also provide an overview of the main DL techniques and approaches aimed at performing disease prediction and prevention and at identifying bio-markers via biomedical data and images. In the second part we describe how we applied DL techniques to perform the segmentation of vessels in the ilio-femoral images. Furthermore, we propose a combination of multi-instance segmentation network and optical flow to solve the multiinstance segmentation and detection tasks in endoscopic images. In the third part a combination of data reduction and data visualization techniques is proposed for the reduction of clinical and omics data and their visualization into images, with the aim of performing DL-based classification. Furthermore, we present a ML-based approach to develop a risk model for class prediction from high-dimensional gene expression data, for the purpose of identifying a subset of genes that may influence the survival rate of specific patients. Eventually, in the fourth part we provide a study on the behaviour of AI-based systems during classification tasks, such as image-based disease classification, which is a widely studied topic in the recent years; more in detail, we show how DL-based systems can be studied with the aim of identifying the most relevant elements involved in the training processes and validating the network’s decisions, and possibly the clinical treatment and recommendation.