HAVOC: Small-scale histomic mapping of cancer biodiversity across large tissue distances using deep neural networks

Abstract

Intratumoral heterogeneity can wreak havoc on current precision medicine strategies because of challenges in sufficient sampling of geographically separated areas of biodiversity distributed across centimeter-scale tumor distances. To address this gap, we developed a deep learning pipeline that leverages histomorphologic fingerprints of tissue to create “Histomic Atlases of Variation Of Cancers” (HAVOC). Using a number of objective molecular readouts, we demonstrate that HAVOC can define regional cancer boundaries with distinct biology. Using larger tumor specimens, we show that HAVOC can map biodiversity even across multiple tissue sections. By guiding profiling of 19 partitions across six high-grade gliomas, HAVOC revealed that distinct differentiation states can often coexist and be regionally distributed within these tumors. Last, to highlight generalizability, we benchmark HAVOC on additional tumor types. Together, we establish HAVOC as a versatile tool to generate small-scale maps of tissue heterogeneity and guide regional deployment of molecular resources to relevant biodiverse niches.

PHARAOH: A collaborative crowdsourcing platform for PHenotyping And Regional Analysis Of Histology

“…By clustering associated image patches using these signatures, we found that we could organize WSI, from a wide diversity of tissue types, into regional partitions that show relative cytoarchitectural uniformity 6 ( Fig 1Ai ). Importantly, these proposed tissue regions correlate with expert annotations, immunohistochemical readouts and even subtle intra-tumoral differences in molecular profiles 7 . For image annotation, this DLFV-based clustering approach therefore has the potential to serve as the foundation of an active weakly-supervised learning approach in which a system only needs to query experts for sparse cluster-level labeling of grouped image patches for custom model training 3,4 .…”

“…Tissue partitions are generated in PHARAOH using an unsupervised image feature-based clustering workflow implemented in python (https://pypi.org/project/havoc-clustering/), as previously described 6,7 . Briefly, WSIs are first tiled into individual 0.066-0.27 mm 2 image patches (patch width: 129 μm (256 pixels) to 258 μm (512 pixels)) respectively.…”

“…Here, we use pathologic, molecular and clinical outcome data to assess if human annotations applied to PHARAOH-proposed clusters, each comprising ∼10 2 -10 3 related image patches, can be used to fine-tune CNNs for custom applications in computational pathology. To organize histological information spanning entire tissue sections into meaningful partitions, PHARAOH divides submitted WSIs into non-overlapping image patches (0.017-0.07 mm 2 ; user-defined size) and uses their individual DLFV signatures to organize tiles a empirically optimized solution of nine image clusters 7 ( Fig 1B-C ). Importantly, when tile partitions are projected back onto the WSI, clusters often show non-random spatial distributions that closely mirror morphologic regions of uniformity that can be easily recognized and intuitively labeled by pathologists ( Fig 1D ).…”