Obtaining Spatially Resolved Tumor Purity Maps Using Deep Multiple Instance Learning In A Pan-cancer Study

Abstract

Tumor purity is the proportion of cancer cells in the tumor tissue. An accurate tumor purity estimation is crucial for accurate pathologic evaluation and for sample selection to minimize normal cell contamination in high throughput genomic analysis. We developed a novel deep multiple instance learning model predicting tumor purity from H&E stained digital histopathology slides. Our model successfully predicted tumor purity from slides of fresh-frozen sections in eight different TCGA cohorts and formalin-fixed paraffin-embedded sections in a local Singapore cohort. The predictions were highly consistent with genomic tumor purity values, which were inferred from genomic data and accepted as the golden standard. Besides, we obtained spatially resolved tumor purity maps and showed that tumor purity varies spatially within a sample. Our analyses on tumor purity maps also suggested that pathologists might have chosen high tumor content regions inside the slides during tumor purity estimation in the TCGA cohorts, which resulted in higher values than genomic tumor purity values. In short, our model can be utilized for high throughput sample selection for genomic analysis, which will help reduce pathologists’ workload and decrease inter-observer variability. Moreover, spatial tumor purity maps can help better understand the tumor microenvironment as a key determinant in tumor formation and therapeutic response.

Artificial intelligence-augmented histopathologic review using image analysis to optimize DNA yield from formalin-fixed paraffin-embedded slides

“…These systems can be combined with digital slide marking (digitally guided macrodissection), enabling integration with computer vision models for tumor enrichment 19,20 . Several computer vision systems have been recently developed with the goal of estimating tumor-rich dissection areas from histopathology slides to meet tumor purity input requirements for molecular testing 6,[21][22][23][24] . However, no recommendation systems exist for estimating tissue quantity for minimum DNA input requirements, and thus even automated dissection systems rely on a pathologist to determine how many slides should be scraped.…”

Weakly-supervised deep learning models in computational pathology

“…Tumour purity estimates are not only reflective of the tumour microenvironment (TME), but may have clinical significance in prognosis and therapeutic response. 6 , 7 In addition to the challenges of inter-observer variability between pathologist scores, these scores have been found to correlate poorly with molecular tumour purity values generated from genomic and gene expression data. 6 These latter methodologies are themselves time-consuming and expensive that while enabling precision oncology nevertheless have limitations in their utility for traditional diagnostic methods.…”

“… 6 , 7 In addition to the challenges of inter-observer variability between pathologist scores, these scores have been found to correlate poorly with molecular tumour purity values generated from genomic and gene expression data. 6 These latter methodologies are themselves time-consuming and expensive that while enabling precision oncology nevertheless have limitations in their utility for traditional diagnostic methods. In this weakly-supervised approach, Brendel et al, employed an attention based, multi-task, multiple-instance learning (MIL) model to learn weight features for ROIs within a slide as well as feature representation that can vie with pathologist-derived estimates of tumour purity, exceeding accuracy of previous supervised learning approaches.…”

“… 4 These spatial maps are necessary for downstream applications including sequencing or proteomics analysis from tissue sections for precision medicine or exploratory research that can assist in understanding the key cell types, including immune cell influence, in the TME that can describe tumour progression. 6 Moreover, as targeted therapies emerge confirming sufficient tumour purity for extraction for subsequent molecular testing is essential to preventing false negative results. 3 Model validation on formalin-fixed paraffin-wax embedded tissue sections, 6 the mainstay of resource-poor countries, is also required given concerns regarding model robustness when taking into account variance in preservation methods on tissue architecture and tissue processing.…”

“… 6 Moreover, as targeted therapies emerge confirming sufficient tumour purity for extraction for subsequent molecular testing is essential to preventing false negative results. 3 Model validation on formalin-fixed paraffin-wax embedded tissue sections, 6 the mainstay of resource-poor countries, is also required given concerns regarding model robustness when taking into account variance in preservation methods on tissue architecture and tissue processing. Nevertheless, by being able to accurately predict tumour purity, such models could assist in improving understanding tumour progression and clinical outcomes without sequencing, particularly in resource-limited settings.…”