Radiomics of gastrointestinal stromal tumors; risk classification based on computed tomography images – a pilot study

Date 28 September 2019
Event ESMO 2019 Congress
Session Poster Display session 1
Topics GIST
Presenter Milea Timbergen
Citation Annals of Oncology (2019) 30 (suppl_5): v683-v709. 10.1093/annonc/mdz283
Authors M.J. Timbergen1, M.P.A. Starmans2, M. Vos3, M. Renckens4, D.J. Grünhagen5, G.J.L.H. van Leenders6, W.J. Niessen2, C. Verhoef5, S. Sleijfer7, S. Klein2, J.J. Visser4
  • 1Surgical Oncology, Erasmus MC Cancer Institute, 3015 GD - Rotterdam/NL
  • 2Radiology And Nuclear Medicine, Medical Informatics, Erasmus MC, 3015 GD - Rotterdam/NL
  • 3Medical Oncology And Surgical Oncology, Erasmus MC Daniel den Hoed Cancer Center, 3075EA - Rotterdam/NL
  • 4Radiology And Nuclear Medicine, Erasmus MC, 3015 GD - Rotterdam/NL
  • 5Surgical Oncology, Erasmus MC Cancer Institute, Rotterdam/NL
  • 6Pathology, Erasmus MC, 3015 CN - Rotterdam/NL
  • 7Medical Oncology, Erasmus MC Daniel den Hoed Cancer Center, Rotterdam/NL

Abstract

Background

Gastrointestinal stromal tumors (GISTs) are rare mesenchymal tumors of the gastrointestinal (GI) tract. The size, the presence of a c-KIT gene mutation and the mitotic index are used to determine the prognosis and to direct systemic treatment choices. This study evaluated the use of radiomics, a technique which uses algorithms for diagnosing and predicting the c-KIT mutational status and mitotic index of GISTs from medical imaging features.

Methods

A combination of machine learning methods was used to distinguish treatment-naive GISTs from other GI tumors resembling GISTs on imaging, based on clinical and molecular characteristics, as well as imaging features extracted from the contrast-enhanced venous phase computed tomography scans. Evaluation was performed in a 100x random-split cross-validation with 20% of the data for testing.

Results

A total of 242 tumors were used for distinguishing GISTs (n = 123) from non-GISTs (n = 119) including leiomyoma (n = 25), schwannoma (n = 21), gastric carcinoma (n = 25), lymphoma (n = 23) and leiomyosarcoma (n = 25).The non-GISTs were located either gastric (23%) or non-gastric (77%) and the GISTs were located either gastric (63%) or non-gastric (37%). A c-KIT mutation was present in the majority of the GISTs (exon 9 n = 10, exon 11 n = 56, exon 13 n = 2). The dataset originated from 65 different scanners, leading to heterogeneity in the imaging protocols. Imaging feature analyses showed a mean area under the curve (mAUC) of 0.70 (95% confidence interval [CI] 0.63-0.76) for distinguishing GIST from non-GISTs. A mAUC of 0.52 (95% CI 0.32-0.72) was found for predicting all c-KIT mutations, a mAUC of 0.51 (95% CI 0.29-0.72) for predicting a c-KIT exon 9 mutation, and a mAUC of 0.61 (95% CI 0.47-0.74) for predicting a c-KIT exon 11 mutation. The mitotic index was available in 83 patients (≤5/50 high power fields (HPFs) n = 53, 43.1%, >5/50 HPFs n = 33, 26.8%), and showed a mAUC of 0.60 (95% CI 0.47-0.73).

Conclusions

This pilot study showed the potential of radiomics to distinguish GIST from other GI tumors, but no potential in predicting c-KIT mutational status and mitotic index of GISTs. Further optimization and validation of the radiomics model is required to incorporate radiomics in the diagnostic routine of GISTs.

Clinical trial identification

Editorial acknowledgement

M.J. Timbergen and M.P.A. Starmans contributed equally to this study.

Legal entity responsible for the study

The authors.

Funding

The Netherlands Organization for Scientific Research.

Disclosure

All authors have declared no conflicts of interest.