Evaluating human-AI interaction in the detection of lung, breast and colorectal cancers: a systematic scoping literature review

Davies, Sarah, Miller, Paul K. ORCID: https://orcid.org/0000-0002-5611-1354 and Donovan, Tim ORCID: https://orcid.org/0000-0003-4112-861X (2024) Evaluating human-AI interaction in the detection of lung, breast and colorectal cancers: a systematic scoping literature review. In: UK Imaging and Oncology Congress 2024 (UKIO 2024): Vision and values: Putting people first, 10-12 June 2024, Liverpool ACC, UK. (Unpublished)

Preview

PDF (poster abstract) - Accepted Version Available under License CC BY-NC Download (74kB) | Preview

Abstract

Background: In forthcoming years, Artificial Intelligence (AI) will become an increasingly integral part of medical imaging practice, and automated image interpretation, including Computer-Aided Detection (CADe), represents a particularly significant application of AI within this domain. It is, therefore, imperative that diagnostic radiographers develop an understanding of the role, value and potential pitfalls of using AI in medical image interpretation. Consequently, this paper reports findings from a systematic scoping literature review of studies addressing human-AI interaction in the detection of lung, breast and colorectal cancers.

Method: The Cochrane Central Register of Controlled Trials, PubMed and ScienceDirect were used to locate CADe observer performance studies in three reading paradigms: second-reader, concurrent and interactive. Conversant experimental studies were also retrieved. Key findings from a total of N=46 articles were extracted and organised according to reading paradigm and imaging modality.

Results: Aggregated evidence indicated that concurrent and second-reader CADe systems improved observer sensitivity, particularly among less experienced observers, but could also deleteriously affect attention allocation. Interactive CADe systems demonstrated a weaker impact upon observer sensitivity, though a smaller number of studies was available for analysis. Evidence also indicated that observer performance with CADe could be influenced by prompt design, level of participant expertise, lesion size and number of CAD false positives.

Conclusions: Concurrent and second-reader paradigms have equivalent potential to improve observer performance, though concurrent CADe is more time-efficient; an important consideration in practical clinical terms. Further research is needed to comprehensively understand the value of interactive CADe, particularly in volumetric imaging.

Item Type:	Conference or Workshop Item (Poster)
Departments:	Institute of Health > Medical Sciences
Depositing User:	Paul Miller
Date Deposited:	24 Mar 2024 12:05
Last Modified:	24 Mar 2024 12:15
URI:	https://insight.cumbria.ac.uk/id/eprint/7581

Downloads