A deep learning (DL) model can be used to predict patient outcomes and mutations using hematoxylin, eosin, and saffron (HES)-stained whole slide images from patients with gastrointestinal stromal tumor (GIST), according to a study published in NPJ Precision Oncology.

The DL model that was trained using HES images performed better at predicting relapse-free survival than using the Miettinen risk criteria, which utilize mitotic count, tumor location, and tumor size (C-index, .81 vs .76; std for both .04). Combining the Miettinen risk criteria into the DL model (Deep Miettinen) yielded a higher but not significant improvement (C-index, .83).

The DL model was able to distinguish between mutations in the PDGFRA, KIT, and other genes (wild-type) at the gene level using the HES images (macro-area under the curve (AUC), .74). Performance was better using tumors from the stomach.

Continue Reading

Further stratification of mutations into PDGFRA exon 18, KIT exon 9, KIT exon 11, other PDGFRA or KIT mutations, and wild type yielded a similar but slightly lower AUC of .69 on the test dataset.

Read more about GIST prognosis

The DL model was also able to make codon-level predictions of PDGFRA exon 18 D842V mutation with an AUC of .9 and KIT codons 557/558 deletion (del-inc 557/558) mutations with an AUC of .87. The PDGFRA exon 18 D842V mutation is generally associated with imatinib resistance while the KIT deletion mutation is associated with a worse prognosis compared to other KIT exon 11 mutations.

The DL mutation models were also better at mutation prediction than a model based on cell morphology (spindle cells, epithelioid cells, or mixed), indicating that the DL model incorporates additional morphological features to distinguish mutational status.

“This study shows that DL could help to predict the risk of progression in localized GIST but needs improvements to be used in the clinical management of patients. DL seems more robust for identifying somatic mutations directly from HES whole slide images of patients with GIST,” the authors wrote.

The authors highlighted that the DL methodology could help speed up the therapeutic decision-making process by predicting PDGFRA exon 18 D842V mutations in intermediate-risk Miettinen patients compared to high-risk Miettinent patients. The DL model could also be used to help identify novel histological features for the prediction of relapses.


Fu Y, Karanian M, Perret R, et al. Deep learning predicts patients outcome and mutations from digitized histology slides in gastrointestinal stromal tumor. NPJ Precis Oncol. 2023;7(1):71. doi:10.1038/s41698-023-00421-9