Project duration:
Nov 2019 - March 2024
Explainable deep learning
Explainable deep learning for diagnostic decisions in healthcare
Project summary:
In a typical medical scenario, the highly skilled and trained medical doctor treats a patient by looking at the clinical symptoms and the medical tests of the patient like diagnosis imaging, genetic testing and electrodiagnosis. The medical data is vast and heterogeneous e.g. textual data, images, electrophysiological signals. With the advent of the era of digitization in healthcare throughout hospitals and general practitioners, the collection of easily accessible medical data is increasing everyday. Articial Intelligence (AI) techniques can help in extracting hidden information and patterns from these medical data that is hard or too time-intensive for humans to uncover with limited time and resources. Imagine, there is a system, which has access to various tests of a patient and also information of patients with similar medical conditions (both rare and common ones) and their outcome. Given limited time and resources of a doctor, it will be very benecial for doctors to have access to a single AI system, which can process these various types of information and assist the doctors to arrive at a faster and more accurate decision making, for example, diagnosis of a disease, deciding on treatments etc. Some of the ways in which the doctors can benet from an AI system include the opportunity to spend the saved time with their patients and getting easier access to recent results about similar medical cases.
Popular AI techniques like machine learning are being used to process medical data, for example, supervised machine learning algorithms like support vector machine, random forest, logistic regression and neural network are being applied to major medical conditions like cancer, neurology and cardiology [10]. However, the traditional machine learning algorithms depend on expert dened feature representation of medical conditions and sometimes, fail to discover novel patterns hidden in the medical data. On the other hand, deep learning, an extension of the traditional neural network, is a current machine learning technique automatically capable of learning abstract feature representation from data and has the capability of handling large, complex and multimodal data. Deep learning is showing great potential in many domains - healthcare being one of them [13], with applications ranging from diagnosis of congenital cataract disease [11], skin cancer from clinical images [7], discovering disease onset from longitudinal patient data [17], predicting eye disease by Google DeepMind [1] etc. Also, deep learning has been found to outperform human experts, for example, in identifying cervical precancer [8] and in predicting pneumonia from chest x-rays [16].
Project leader:
Funding:
