An Expert-Driven Machine Learning for Medical Imaging-Based Diagnosis of Lesions

Le résumé fourni par la source

This paper describes an expert-driven, instead of data-driven, machine learning (EdML) to characterize in vivo tissues for medical imaging-based diagnosis of lesions. The presented EdML mimics experts’ performance by incorporating experts’ prior knowledge into analysis of data to accomplish a task that the data were acquired for. The experts’ prior knowledge for the task of lesion diagnosis include: (1) lesions are the outcomes of in vivo tissues’ biological changes from normal to abnormal; (2) lesion diagnosis is based on pathological conditions (malignant vs. benign) of the tissue biological changes; (3) lesion pathological conditions are described by tissue pathology and thus tissue pathology must be used as the ML training labels for the task; and (4) what tissue biological information is encoded inside the data by the imaging device. This study tested the EdML using low dose computed tomography (LdCT) imaging modality for the objective of early lung cancer screening. The challenge to achieve the objective is the diagnosis of those screening-detected pulmonary nodules, which are interpreted as indeterminate lesions (IDLs) by medical experts in the field. Current data-driven ML (DdML) cannot perform better than the medical experts. Because of the difficulty in diagnosing the IDLs, both experts and current DdML reached AUC (area under receiver characteristic curve) of 0.89. Two IDL datasets (sample sizes of 68 and 114) with their tissue pathological reports as ML labels were used to quantify the performances of EdML with comparison to DdML with AUCs of 0.90 s vs. 0.70 s. These experimental outcomes demonstrated an immense potential of EdML to overcome the challenge to achieve the objective of LdCT screening of early lung cancer.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.