An algorithm developed by deep learning from digitalized hematoxylin and eosin-stained whole tumor slide images outperformed classical Miettinen relapse risks prediction in patients with gastrointestinal stromal tumors (GIST). A second algorithm predicted mutations with high accuracy.


Gastrointestinal stromal tumors (GIST), the most frequent mesenchymal tumor of the GI tract, shows a variable clinical behavior ranging from benign to malignant. Risk assessment according to the AFIP/Miettinen classification (high, intermediate, or low risk of relapse) and mutational profiling are major tools for patient management. AFIP/Miettinen classification includes size of the tumor, localization, and mitotic count. However, Miettinen classification comes with subjectivity (mitotic count) and is time-consuming. In addition, mutational profiling is costly, time-consuming, and not yet available in all countries (or centers). Therefore, Dr. Raul Perret (Institut Bergonié, France) and colleagues evaluated the efficacy of deep learning models to predict relapse-free survival in patients with GIST and to predict mutational profiles.

Both models were based on histology (ie, digitalized hematoxylin and eosin-stained whole tumor slide images). The researchers trained the relapse-predicting model using whole tumor slide images data from 305 patients from one institute and validated the model using data from 286 patients from a different center. Both cohorts had similar distribution of GIST types (localization, TKI treatment). Likewise, training of the model for prediction of mutation profile was performed using data from 1,233 patients from the one institute and validation on data from 238 patients from the other center.

The algorithm for relapse prediction proved to outperform prediction based on Miettinen classification (C-index 0.81 vs 0.76). Combining deep learning with tumor location and size (Deep Miettinen) further improved C-index to 0.83. Deep Miettinen was able to stratify patients as high or low risk for relapse-free survival. In addition, the algorithm was able to dichotomize patients characterized as “high risk for relapse” according to classical Miettinen into two groups: high versus low risk. Likewise, the algorithm was able to dichotomize classical “intermediate risk” patients into a high risk or low risk group. Histological features associated by the algorithm with “high risk” are mitosis, marked nuclear atypia, high cellular density, epithelioid cell component, necrosis, and hemorrhage. Histological features associated by the algorithm with “low risk” are cytoplasmic vacuolization, low cellular density, collagenous stroma, mild nuclear atypia, and spindle cell component.

The algorithm for prediction of the presence of mutations also performed well. The area under the curve (AUC) for predicting KIT-mutations was 0.80 in the training cohort and 0.85 in the validation cohort. AUC for predicting PDGFRA-mutations was 0.92 in both cohorts. More specifically, AUC for predicting PDGFRA exon 18 D824V mutation was 0.87 in both cohorts and for predicting KIT exon 11 del 557–558 mutation was 0.69 in the training cohort and 0.76 in the validation cohort. Histological features associated by the algorithm with KIT exon 11 del 557–558 were mitotic activity and nuclear hyperchromasia; histological features associated by the algorithm with PDGFRA exon 18 D824V mutation were epithelioid or mixed cell morphology, cytoplasmic vacuolization, myxoid stroma, and lymphoid infiltrate.

“These results show that the deep learning model outperforms Miettinen in predicting relapse-free survival in localised untreated GIST,” concluded Dr. Perret. “Using Deep Miettinen is possible to stratify existing risk groups in Miettinen. In addition, the deep learning model predicts mutations with high accuracy. Both models identified histological features associated with risk of relapse and mutational profile, respectively. However, further validation of the models is needed.”

 

Copyright ©2022 Medicom Medical Publishers