An Investigation of Global and Local Radiomic Features for Customized Self-Assessment Mammographic Test Sets for Radiologists in China in Comparison with Those in Australia

Abstract

Self-assessment test sets have demonstrated being effective tools to improve radiologists’ diagnostic skills through immediate error feedback. Current sets use a one-size-fits-all approach in selecting challenging cases, overlooking cohort-specific weaknesses. This thesis assessed feasibility of using a comprehensive set of handcrafted global radiomic features (Stage 1, Chapter 3) as well as handcrafted (Stage 2, Chapter 4) and deep-learning based (Stage 3, Chapter 5) local radiomic features to identify challenging mammographic cases for Chinese and Australian radiologists. In the first stage, global handcrafted radiomic features and Random Forest models analyzed mammography datasets involving 36 radiologists from China and Australia independently assessing 60 dense mammographic cases. The results were used to build and evaluate models’ performance in case difficulty prediction. The second stage focused on local handcrafted radiomic features, utilizing the same dataset but extracting features from error-related local mammographic areas to analyze features linked to diagnostic errors. The final stage introduced deep learning, specifically Convolutional Neural Network (CNN), using an additional test set and radiologists’ readings to identify features linked to false positive errors. Stage 1 found that global radiomic features effectively detected false positive and false negative errors. Notably, Australian radiologists showed less predictable errors than their Chinese counterparts. Feature normalization did not improve model performance. In Stage 2, the model showed varying success rates in predicting false positives and false negatives among the two cohorts, with specific mammographic regions more prone to errors. In Stage 3, the transferred ResNet-50 architecture performed the best for both cohorts. In conclusion, the thesis affirmed the importance of radiomic features in improving curation of cohort-specific self-assessment mammography test sets.