Assessing the completeness of reporting in imaging studies using artificial neural network models for cancer
Tayser Zoubi1, Samuel J White2, Shubham Tiwari1
1College of Medicine and Public Health, Flinders University, Adelaide, South Australia, Australia.
Background:
Developments in artificial neural networks (ANNs) offer significant promise for cancer screening and risk prediction, with the potential to improving patient outcomes. Ensuring complete and transparent reporting of study methodologies is important for ensuring model reproducibility. This review aims to evaluate the completeness of reporting of imaging studies that utilise ANN models for cancer screening and characterisation.
Methods:
Studies employing novel or substantively modified ANN architectures in the screening, diagnosis, or classification of neoplasms were analysed. Completeness of reporting was assessed using the TRIPOD + AI checklist.
Results:
The search strategy identified 1379 studies, of which 100 were included in this analysis. Across these studies, the reporting adherence to the TRIPOD-AI checklist was moderate with a mean of 69% (range 65-75%). Items relating to data processing practices, missing data management strategies, and external validation methods were infrequently reported (reported by less than 33% of studies). There was considerable heterogeneity in the reporting of model diagnostic performance metrics. Sensitivity, specificity, and accuracy were the most reported evaluation metrics, featured in 61%, 34%, and 45% of included studies, respectively. Clinically relevant outcome measures were infrequently reported.
Conclusion:
The reproducibility of radiological studies using ANN-based models to screen or characterise cancer is limited by suboptimal reporting practices. Potential measures to support more complete reporting include making adherence to appropriate reporting guidelines a condition of manuscript submission and mandating code and data sharing practices. Furthermore, stronger emphasis on reporting clinically relevant outcome measures (as opposed to just statistical measures of model performance) would greatly support decision-making with respect to implementation of study findings into clinical practice.
