Session Item

To propose a novel method to evaluate the stability of radiomic features (RFs) extracted from multiparametric (mp)MRI sequences to variations in segmentation using in-silico contour generation. We applied the process to the whole prostate region. The stable RFs identified are then used to extract a robust subset of radiomic features. 

To evaluate the stability of RFs to variations in segmentation, we simulated various under- and/or over-segmentation scenarios commonly seen in clinical practice using the principle of data augmentation in deep learning. The overall workflow is illustrated in Figure 1.a. The manual delineation by a single radiologist was perturbed using geometric transformations such as rotation, scaling, and shifting to generate the synthetic segmentations. We considered 3 categories of augmentations (in-plane, out-plane, and in&out-plane) and 2 types of biases (random and systematic) as part of the study (see Fig. 2 for definitions). T2w, ADC, and DCE-SUB sequences with delineated whole prostate were considered as part of the investigation. We extracted 1595 RFs (using Pyradiomics) from each image sequence-segmentation pair. We used a wide range of filters such as LoG, wavelet, squared, square root, logarithm, and exponential filters to see the impact of various filtering strategies on the stability of RFs. 

Fig. 1.b summarizes the datasets used in this study. We used the internal population to identify the stable RFs, while we used the external set to select the robust subset of RFs (by considering the overlap). We chose the intraclass correlation coefficient, ICC(1,1), to assess the stability of RFs, and we considered a feature as stable if the lower bound associated with the 95% CI of the ICC estimate had a value ≥ 0.90. 

Here, the results are simplified to present only the best filter (or non-filter) that led to the maximum number of stable and robust RFs (different filters for different RFs families). Table 2 summarizes the percentage of best-filtered stable & robust RFs derived from T2w, ADC, and DCE-SUB maps.  

Variations in segmentation impact the stability of RFs. We propose a novel method that can be easily integrated into any pipeline, which can be used to identify RFs stable to such variability. Since this is a fully automated method, it avoids the need for multiple radiologists to annotate the dataset. The robust subset of features identified can be considered for further modeling or analysis.