A mutation-based radiomics signature predicts response to imatinib in Gastrointestinal Stromal Tumors (GIST)

Abstract

Objectives: To develop a mutation-based radiomics signature to predict response to imatinib in Gastrointestinal Stromal Tumors (GISTs).

Methods: Eighty-two patients with GIST were enrolled in this retrospective study, including 52 patients from one center that were used to develop the model, and 30 patients from a second center to validate it. Reference standard was the mutational status of tyrosine-protein kinase (KIT) and platelet-derived growth factor α (PDGFRA). Patients were dichotomized in imatinib sensitive (group 0 - mutation in KIT or PDGFRA, different from exon 18-D842V), and imatinib non-responsive (group 1 - PDGFRA exon 18-D842V mutation or absence of mutation in KIT/PDGFRA). Initially, 107 texture features were extracted from the tumor masks of baseline computed tomography scans. Different machine learning methods were then implemented to select the best combination of features for the development of the radiomics signature.

Results: The best performance was obtained with the 5 features selected by the ANOVA model and the Bayes classifier, using a threshold of 0.36. With this setting the radiomics signature had an accuracy and precision for sensitive patients of 82 % (95 % CI:60-95) and 90 % (95 % CI:73-97), respectively. Conversely, a precision of 80 % (95 % CI:34-97) was obtained in non-responsive patients using a threshold of 0.9. Indeed, with the latter setting 4 patients out of 5 were correctly predicted as non-responders.

Conclusions: The results are a first step towards using radiomics to improve the management of patients with GIST, especially when tumor tissue is unavailable for molecular analysis or when molecular profiling is inconclusive.

Keywords: Artificial intelligence; GIST; Mutational status; Radiomics; Response to therapy.

Fig. 1

Flowchart of the study.

Fig. 2

Balanced accuracy and delta obtained on the construction set for all combinations of features selection and classifier. Delta represents the difference between the accuracy in detecting imatinib sensitive and imatinib non-responsive patients.

Fig. 3

Waterfall plot of the prediction probabilities of the Bayes Classifier. The radiomics score of 0.36 is normalized to 0. Negative radiomics score represent the probability of being imatinib sensitive while positive radiomics score represent the probability of being non-responsive. Therefore, green bars having a positive section represent patients incorrectly classified as non-responsive, while red bars having a negative radiomics score are patients incorrectly classified as sensitive. The black line represents the threshold that maximizes accuracy on sensitive patients resulting in a lower number of patients correctly classified as non-responsive (reduced accuracy), but an higher number of correctly classified imatinib non-responsive patients over the total number of patients classified as non-responsive (increase of precision for imatinib non-responsive patients).

Fig. 4

Weights of each selected feature on the Bayes Classifier.

Fig. 5

four portal phase CT scans of different patients. GIST with KIT exon 11 mutation (imatinib sensitive) shows a homogeneous and hypodense appearance (Fig. 5A, white arrows), while GIST with WT status (non-responsive) shows an heterogeneous appearance with different regions of higher contrast enhancement (Fig. 5B, white arrows). Both GISTs were correctly predicted by our radiomics model. Fig. 5C (white arrows) shows a KIT exon 11 GIST (imatinib sensitive) and it represents one of the eight imatinib-sensitive GIST erroneously predicted as non-responsive. At CT scan it shows an heterogeneous enhancement, differently than expected. Conversely, Fig. 5D (white arrows) shows a PDGFRA D842V GIST (non-responsive) where we expected an heterogeneous enhancement, while CT images showed an homogeneous appearance: this example represent one of the eight non-responsive GIST erroneously predicted as imatinib sensitive.