Radiomic Features on Diagnostic Magnetic Resonance Enterography Appear to Predict Patient Outcome Following Treatment of Crohn’s Disease: Preliminary Results
Cheng Lu1, Maneesh Dave2, H. Matthew Cohn2, Prateek Prasanna1, Jeffrey Katz2, Rajmohan Paspulati3, Anant Madabhushi1, and Satish Viswanath1

1Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, 2Digestive Health Institute, University Hospital, Case Western Reserve University, Cleveland, OH, United States, 3Dept of Radiology, University Hospital, Case Western Reserve University, Cleveland, OH, United States

Synopsis

We present initial results of identifying radiomic features (computer-extracted image features from radiography) from baseline diagnostic Magnetic Resonance Enterography (MRE) scans to discriminate patients who will and will not respond to initial therapy for Crohn’s Disease. Feature selection was employed to identify the most discriminatory features, followed by principal component analysis to identify an optimal combination of these features. In a cohort of 11 patients, the radiomic feature combination was able to successfully distinguish between responders and non-responders with only 1 misclassification. Multi-scale oriented gradient (Gabor) features appeared to best capture subtle inflammation-related imaging characteristics on MRE and hence most predictive of patient outcome.

PURPOSE:

We present initial results of identifying radiomic features to discriminate between patients who will and will not respond to initial therapy for Crohn’s Disease, derived from a diagnostic Magnetic Resonance Enterography (MRE) scan. It is reported that up to 40% of Crohn’s patients may not respond to initial therapy, and may instead suffer from increasing severity of small bowel inflammation [1]. This suggests an urgent need to better segregate patients who will most benefit from such therapy during diagnosis. MRE is a high-resolution, non-invasive imaging modality of choice to evaluate Crohn’s disease presence and activity in vivo [2]. However, expert interpretation of MRE remains subjective, causing scores based on such assessment to be highly variable [3]. Further, the structural complexity of the small bowel makes expert interpretation difficult. Radiomic features have been shown to capture subtle imaging features of disease appearance [4], by quantifying higher order information that may not be discernible upon visual inspection. Our hypothesis is that such advanced imaging features can capture information indicative of patient outcome to treatment, which is not based on disease severity alone.

METHODS:

A cohort of 30 patients were considered in this study, all of whom had been imaged via MRE on a Siemens scanner, as part of the diagnostic clinical protocol for Crohn’s Disease. Scans from only 11 patients could be utilized due to either poor quality imaging, fundamental scanner/protocol differences, or colonic inflammation (instead of the small bowel). The true fast imaging sequence with steady precession and fat suppression (TRUFIFS) was used. Annotations of Crohn’s extent on the TRUFIFS images were obtained from an expert radiologist, based on standard guidelines [5], and used as the basis for evaluation of radiomic features. Our radiomic feature analysis comprised the following steps: (i) bias field correction to remove acquisition-related intensity variation across the image volume, (ii) radiomic feature extraction of 95 pixel-wise features, comprising 1st and 2nd order statistics, oriented wavelets (Gabor), and local neighborhood-based texture energy (Laws) [4], (iii) feature pruning, to retain the top 10 radiomics features whose distributions showed the most separation between the 2 patient groups, as quantified by the Bhattacharyya distance [6], and (iv) Principal Component Analysis to project and visualize patient-wise radiomic features (median, variance, skewness, and kurtosis over all the pixels within each annotation) into a 3-dimensional Eigen space.

RESULTS AND DISCUSSION:

Of the 11 patients in our cohort, 5 patients responded and were stable to initial treatment for Crohn’s while the remaining 6 did not respond and therapy ultimately had to be escalated in these patients; as determined via patient reports. The 3D scatter plot of the 11 patients in PCA-reduced Eigen space (Figure 1) suggests good separability between the 2 groups, with only 1 misclassification (90.9% accuracy). While the original MRE images of the 2 patients do not indicate marked visual differences in appearance, the radiomic feature heatmaps demonstrate markedly different signatures within manually annotated regions of interest in the small bowel between patients that did and did not respond to treatment (shown in Figure1-(b), (d)). The most discriminating radiomic features were found to be multi-scale oriented Gabor features, which capture a combined response for both inflamed bowel wall regions and lumen regions within disease ROIs, possibly capturing disease activity and not severity alone.

CONCLUDING REMARKS

We presented preliminary results of identifying radiomic features from a baseline MRE scan to segregate patients with Crohn’s Disease who responded favorably to initial therapy from those who did not respond and needed an escalation of therapy. Gabor features which attempted to capture combined multi-scale gradient response of inflamed bowel wall and lumen were found to be most predictive. This work could potentially pave the way for optimal patient and therapy management for Crohns disease, as well as possibly reducing the need for invasive procedures to assess treatment response.

Acknowledgements

No acknowledgement found.

References

[1] Lapidus A, et al. Gastroenterology. 1998;114(6): 1151–1160. [2] Griffin N, et al., Insights Imaging. 2012; 3(3): 251–263. [3] Spilseth BD, et al. In: RSNA. 2013. [4] Viswanath S et al. Neurocomputing. 2014 Nov; 144(20): 13–23. [5] Mowat C, et al., Gut. 2011 May; 60(5): 571–607. [6] Reyes-Aldasoro, C.C., et al. Pattern Recognition, 2006, Vol. 39, Issue 5, pp. 812–826.

Figures

Figure1. Left: 3D scatter plot via the output of PCA: squares and stars indicate patients who responded to and did not respond to treatment, respectively. The ellipse shows the highly separable clusters. Right: first row and second row show the MRE image with manually annotated ROIs within the small bowel (a)(c), Gabor feature heatmap (b)(d) of patients who did and did not respond to treatment, respectively. Blue indicates low feature expression, whereas red reflects high feature expression.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
2965