Deep Learning and Recommender Systems for Radiography Image Assessment 

Henry Blanco is a Joint PhD student at the “Vrije Universiteit Brussel” (VUB) and at the “University of Oriente” in Cuba, where he has got a bachelor and a master degree in Computer Science. He is currently doing PhD research activities at the department of Electronics and Informatics (ETRO) at the VUB. His research interests are mainly focused on computer aided radiology diagnosis based on the combined use of deep learning and recommender system models.

"Adoption of AI-based medical decision making systems depends on, not only accurate or precise diagnosis models, but on explainable ones"

Current radiology CAD systems, based on convolutional neural networks (CNN) models, are mainly characterized by its high accuracy on classification or detection tasks. The fact that these systems fail or have a poor capacity to explain their decisions, has negatively impacted on its adoption in true clinical environments. These issues specifically affects trust, confidence and decisions robustness among radiologists and its introduction into realistic radiology workflow, as well as the automatic generation of comprehensive radiology reports.

Since “conversational recommender systems” are useful tools, during a series of user-system interactions, to be properly informed on decision making tasks and that “convolutional neural networks” currently represent the state of the art of AI models on classification tasks, we conjecture that the combined use of CNN models using a hierarchy-based conversational recommendation approach on medical image classification tasks, its a promising approach to tackle the problem of accurate and explainable classification deep learning models. Moreover, this kind of approach encourages the automatic creation of radiology reports, which indeed represents the ultimate goal of a medical imaging diagnosis task.

Other Links