Association of diagnostic errors in interpreting screening mammograms with image and reader characteristics

Abstract

Aims

This thesis investigated factors affecting radiologists' performance in mammography interpretation. It examined the association between subjective expert-determined difficulty and objective measures based on reader performance, explored reader-based and image-based influences on interpretation, and used quantised textural analysis to assess how image textures affect reader performance.

Methods

The study used linear mixed modelling, involving 479 Australian radiologists. The initial study involved 165 readers and one test set, while subsequent studies used six test sets and all 479 readers. It used multivariate model analysis with three models based on case-based, reader-based, and mixed characteristics. The final study utilised radiomics analysis.

Results

Findings demonstrated that expert-predicted difficulty in breast cancer cases correlated with breast density but did not align with actual difficulty. In normal cases, image-based characteristics significantly influenced expert-determined difficulty, while reader-based factors like weekly case interpretation and specialisation mattered. In cancer cases, image-based factors weren't significant. The mixed model analysis indicated that increasing expert-predicted difficulty reduced reader performance. The number of weekly interpreted cases also significantly affected performance. Using quantised features enhanced cancer detection compared to non-quantised textural features, with stellate masses outperforming spiculated masses. Consistently, higher expert-predicted difficulty reduced performance, while more weekly cases improved it.

Conclusion

The thesis highlights the importance of expert-predicted difficulty, weekly case volume, and reader- and image-based characteristics in mammography interpretation. It emphasises the need for continued exploration of texture-based radiomic analysis, contributing to improved screening and diagnostic practices in breast cancer detection and management.