AI for Medical Imaging: from the research laboratories to the clinical practice

Motivations

Nowadays, Artificial intelligence (AI) is being demonstrated as one of the most promising solutions to many real-world problems, including various important medical tasks such as tissue/organ segmentation, abnormality detection, disease classification, etc., and many efforts are being made to include AI-based solutions in the clinical practice. 

This is largely thanks to the progress in machine learning and deep neural networks, the availability of cloud computing infrastructure, the large amount of data available to us, and the increased interest of clinicians to pursue research in this field. In fact, it is evident that once established in the hospital’s workflows, AI-based solutions will introduce key opportunities including precision medicine, personalized therapies, and computer-aided clinical decision systems. 

Despite its benefits and potential applications, often proclaimed by the scientific community, deployment of AI algorithms in the clinical routine remains rare due to specific challenges.

This special track aims to gather efforts to understand, limit, and/or address the challenges that are encountered when deploying AI-based solutions to the clinical workflow, including but not limited to:

1. Data governance: it has been shown that AI model performance increases based on the volume of training data available. Even if medical imaging produces an incredible volume of data continuously, the barriers to data access and harnessing prevent its maximum potential. For this reason, how is the data stored and secured, how long is it kept, how is the quality of the data maintained, who has access to the database, and what policies and protocols are in place to ensure the privacy of patient data are fundamental requirements that need to be addressed. 
2. AI robustness: how an AI model may fail to generalize in the presence of a mismatch between the development dataset and the clinical dataset, e.g. dissimilar cohorts, inconsistent acquisition protocols, variability of data across the clinical sites, or annotation policy.
3. Annotated data scarcity: while acquiring expert annotations is expensive, weakly annotated or unlabelled data are often available and might be exploited for learning models.
4. Data uncertainty: how the training of AI models could be ineffective due to the uncertainty (related to intra- and inter-observer variability) affecting the data labels.
5. Data multimodality: how AI models have to fuse different modalities, which modalities have to be fused, and how to manage the possibility of unavailable modalities in a clinical workflow.
6. Ethics: ethical implications of using AI in the clinical workflow regarding data (ownership and privacy, biases, data annotation, and ground-truth generation), algorithms, trained models, and medical practice.

Recent works have pointed out the above difficulties and the scientific community is showing a growing interest in the area of this special track.

The “AI for Medical Imaging: from research laboratories to the clinical routine” special track provides a forum for the discussion of the impact of AI on medical data and how to bridge the gap between the AI technologies developed in the research laboratories and the requirements that have to be fulfilled for the clinical usage. 

Topics

The topics of interest include but are not limited to:

Important Dates

The special track will take place in parallel with the general conference track. Submission deadlines are reported here.

Submission Guidelines

Please refer to this page.

Special Issue

Available Soon

Organizers

Program Committee