Automatic tumour segmentation in brain images: moving towards clinical implementation

Abstract

The aim of this thesis was to examine and enhance the scientific groundwork for translating deep learning (DL) algorithms for brain tumour segmentation into clinical decision support tools. Paper II describes a scoping review conducted to map the field of automatic brain lesion segmentation on magnetic resonance (MR) images according to a predefined and peer-reviewed study protocol (Paper I). Insufficient preprocessing description was identified as one factor hindering clinical implementation of the reviewed algorithms. A reproducibility and replicability analysis of two algorithms was described in Paper III. The two algorithms and their validation studies were previously assessed as reproducible. In this experimental investigation, the original validation results were reproduced and replicated for one algorithm. Analysing the reasons for failure to reproduce validation of the second algorithm led to a suggested update to a commonly-used reproducibility checklist; the importance of a thorough description of preprocessing was highlighted. In Paper IV, radiologists' perception of DL-generated brain tumour labels in tumour volume growth assessment was examined. Ten radiologists participated in a reading/questionnaire session of 20 MR examination cases. The readers were confident that the label-derived volume change is more accurate than their visual assessment, even when the inter-rater agreement on the label quality was poor. In Paper V, the broad theme of trust in artificial intelligence (AI) in radiology was explored. A semi-structured interview study with twenty-six AI implementation stakeholders was conducted. Four requirements of the implemented tools and procedures were identified that promote trust in AI: reliability, quality control, transparency, and inter-organisational compatibility. The findings indicate that current strategies to validate DL algorithms do not suffice to assess their accuracy in a clinical setting. Despite the recognition from radiologists that DL algorithms can improve the accuracy of tumour volume assessment, implementation strategies require more work and the involvement of multiple stakeholders.