Abstract
A quantitative model to genetically interpret the histology in whole microscopy slide images is desirable to guide downstream immuno-histochemistry, genomics, and precision medicine. We constructed a statistical model that predicts whether or not SPOP is mutated in prostate cancer, given only the digital whole slide after standard hematoxylin and eosin [H&E] staining. Using a TCGA cohort of 177 prostate cancer patients where 20 had mutant SPOP, we trained multiple ensembles of residual networks, accurately distinguishing SPOP mutant from SPOP non-mutant patients (test AUROC=0.74, p=0.0007 Fisher’s Exact Test). We further validated our full metaensemble classifier on an independent test cohort from MSK-IMPACT of 152 patients where 19 had mutant SPOP. Mutants and non-mutants were accurately distinguished despite TCGA slides being frozen sections and MSK-IMPACT slides being formalin-fixed paraffin-embedded sections (AUROC=0.86, p=0.0038). Moreover, we scanned an additional 36 MSK-IMPACT patients having mutant SPOP, trained on this expanded MSK-IMPACT cohort (test AUROC=0.75, p=0.0002), tested on the TCGA cohort (AUROC=0.64, p=0.0306), and again accurately distinguished mutants from non-mutants using the same pipeline. Importantly, our method demonstrates tractable deep learning in this “small data” setting of 20-55 positive examples and quantifies each prediction’s uncertainty with confidence intervals. To our knowledge, this is the first statistical model to predict a genetic mutation in cancer directly from the patient’s digitized H&E-stained whole microscopy slide. Moreover, this is the first time quantitative features learned from patient genetics and histology have been used for content-based image retrieval, finding similar patients for a given patient where the histology appears to share the same genetic driver of disease i.e. SPOP mutation (p=0.0241 Kost’s Method), and finding similar patients for a given patient that does not have have that driver mutation (p=0.0170 Kost’s Method).
Significance Statement This is the first pipeline predicting gene mutation probability in cancer from digitized H&E-stained microscopy slides. To predict whether or not the speckle-type POZ protein [SPOP] gene is mutated in prostate cancer, the pipeline (i) identifies diagnostically salient slide regions, (ii) identifies the salient region having the dominant tumor, and (iii) trains ensembles of binary classifiers that together predict a confidence interval of mutation probability. Through deep learning on small datasets, this enables automated histologic diagnoses based on probabilities of underlying molecular aberrations and finds histologically similar patients by learned genetic-histologic relationships.
Conception, Writing: AJS, TJF. Algorithms, Learning, CBIR: AJS. Analysis: AJS, MAR, TJF. Supervision: MAR, TJF.
