Evaluation of Use of Diagnostic AI for Lung Cancer in Practice

Not Applicable

• Conditions: Lung Cancer
• Interventions: Behavioral: AI-human interaction

• Registration Number: NCT03780582
• Lead Sponsor: Ensemble Group Holdings, LLC

Brief Summary

This study investigates ways of improving radiologists performance of the classification of CT-scans as cancerous or non-cancerous. Participants interact with an AI to classify CT-scans under three different conditions.

Detailed Description

The three conditions are as follows: "probabilistic classification", where the radiologist diagnoses scans using an AI cancer likelihood score; "classification plus detection", where the radiologist see detecting lung nodules in addition to the AI's probabilistic classification score before making her own examination of the CT-scan; and "classification with delayed detection", where the radiologist identifies regions of interest independently of the AI and then sees the AI's detected ROIs.

Study Timeline

• Study Start: 2018-12-14
• Study First Submitted: 2018-12-16
• Study First Submitted QC: 2018-12-17
• Study First Posted: 2018-12-19
• Status Verified: 2019-07
• Last Update Submitted: 2019-07-20
• Last Update Posted: 2019-07-23
• Primary Completion: 2019-12-15
• Study Completion: 2019-12-15

Recruitment & Eligibility

• Status: UNKNOWN
• Sex: All
• Target Recruitment: 15

Inclusion Criteria

• The participant performs radiology screenings professionally

Exclusion Criteria

Study & Design

• Study Type: INTERVENTIONAL
• Study Design: CROSSOVER

Arm && Interventions

Group	Intervention	Description
Probabilistic Classification	AI-human interaction	Radiologists see a "score" from 1-100 that represents the AI's prediction of whether the CT-scan comes from a patient with cancer or not before beginning their analysis of the scan.
Classification Plus Detection	AI-human interaction	Radiologists see a "score" from 1-100 that represents the AI's prediction of whether the CT-scan comes from a patient with cancer or not before beginning their analysis of the scan. They also see ROIs identified by the AI that represent lung nodules.
Classification With Delayed Detection	AI-human interaction	Radiologists see a "score" from 1-100 that represents the AI's prediction of whether the CT-scan comes from a patient with cancer or not before beginning their analysis of the scan. After identifying their own ROIs, the radiologist then can see ROIs identified by the AI that represent lung nodules before making final decisions.

Primary Outcome Measures

Name	Time	Method
Classification accuracy	up to 4 months after initiation of evaluation of the test set	This compares radiologists' classifications with the ground truth in the tested cases.

Secondary Outcome Measures

Name	Time	Method
detection concordance	up to 4 months after initiation of evaluation of the test set	Evaluation of concordance between radiologists in the tested cases in detection of lung nodules \> 4 mm

Trial Locations

Locations (1)

University of Hong Kong; 🇭🇰 Hong Kong, Hong Kong