Global radiomic features for improving the process of curating educational mammographic test sets

Abstract

This thesis explores the value of global mammographic radiomic features (GMRFs) for improving the process of curating educational test sets. It first examines existing literature and then studies the value of GMRFs in predicting global gist signals (radiologists’ first impression about an image), and difficult normal cases of radiologists (widely varying experience) and radiology trainees (limited skills).

Three original research studies with ethics approval were included. 1st study used 4191 mammograms to collect gist signals, from 13 radiologists, which were averaged and given to each image. Images were grouped into high- and low-gist. 130 GMRFs/image were extracted and used to build eight models for describing high-gist images. 2nd study, 361 normal cases interpreted by 537 readers were classed into highly difficult and easy. 34 GMRFs/image were extracted and used to create three models to predict difficult cases. 3rd study grouped 280 normal cases, interpreted by 137 radiology trainees, into hardest and easiest. 34 GMRFs/case were retrieved and used to construct a model for predicting hardest cases. The models were trained and validated using 10-fold cross-validation (1st study) and leave-one-out-cross-validation approach (2nd and 3rd study). Their performances were assessed by the AUC. Important features were identified via exploratory factor analysis (all studies).

The models achieved AUCs of up to 0.84 and 0.71, and 0.75 AUC for predicting high-gist images (1st study), and difficult normal cases of radiologists (2nd study) and radiology trainees (3rd study), respectively. Overall, eleven important features were cluster shade, sum entropy, difference variance, correlation, homogeneity, cluster prominence, standard deviation, skewness, kurtosis, range, and coarseness.

The findings offer new insights into possible appearances of GMRFs underlying radiological errors and difficult normal cases, which can be used to enhance educational mammographic test sets.