A Machine Learning Model Based on Global Mammographic Radiomic Features Can Predict Which Normal Mammographic Cases Radiology Trainees Find Most Difficult

TL;DR

This study shows that machine learning using mammographic features can predict which normal mammograms are hardest for trainees to interpret.

Contribution

A novel machine learning model using global radiomic features to predict difficulty in interpreting normal mammograms for trainees.

Findings

The model achieved an AUC of 0.75 in predicting hardest-to-interpret normal cases.

Cluster prominence and range were the most useful features for the model.

Fifteen GMRFs significantly differed between hardest- and easiest-to-interpret cases.

Abstract

This study aims to investigate whether global mammographic radiomic features (GMRFs) can distinguish hardest- from easiest-to-interpret normal cases for radiology trainees (RTs). Data from 137 RTs were analysed, with each interpreting seven educational self-assessment test sets comprising 60 cases (40 normal and 20 cancer). The study only examined normal cases. Difficulty scores were computed based on the percentage of readers who incorrectly classified each case, leading to their classification as hardest- or easiest-to-interpret based on whether their difficulty scores fell within and above the 75th or within and below the 25th percentile, respectively (resulted in 140 cases in total used). Fifty-nine low-density and 81 high-density cases were identified. Thirty-four GMRFs were extracted for each case. A random forest machine learning model was trained to differentiate between hardest- and easiest-to-interpret normal cases and validated using leave-one-out-cross-validation approach. The model’s performance was evaluated using the area under receiver operating characteristic curve (AUC). Significant features were identified through feature importance analysis. Difference between hardest- and easiest-to-interpret cases among 34 GMRFs and in difficulty level between low- and high-density cases was tested using Kruskal–Wallis. The model achieved AUC = 0.75 with cluster prominence and range emerging as the most useful features. Fifteen GMRFs differed significantly (p < 0.05) between hardest- and easiest-to-interpret cases. Difficulty level among low- vs high-density cases did not differ significantly (p = 0.12). GMRFs can predict hardest-to-interpret normal cases for RTs, underscoring the importance of GMRFs in identifying the most difficult normal cases for RTs and facilitating customised training programmes tailored to trainees’ learning needs.

Figures