Impact of Acquisition Parameters on the Performance of Radiomic Systems

Radiomics has emerged as a promising tool for early diagnosis of cancer using Computerised Tomography (CT) scans. However, a main concern is the reproducibility of results, particularly in cases unseen by the network. Although overfitting can contribute to the lack of reproducibility, a broader approach is necessary to consider the variability of visual features introduced by different acquisition parameters of CT scans. In this paper, we statistically analyse the impact of CT scan acquisition parameters on the performance of radiomic methods based on deep learning. We propose using generalised mixed models with fixed factor critical acquisition parameters to estimate significant differences in odds ratio of method failure. We apply our method to two different visual representation spaces based on pre-trained deep networks, VGG and MobileNet. Results on an own collected database for small lung nodule diagnosis, indicates that VGG is more invariant to acquisition parameters than MobileNet and detects pitch spiral as a critical parameter.