V67: Deep Learning Analysis of Sarcoma Histology Correlates Imaging Features with Clinicopathologic Characteristics

Sarcomas are a diverse group of tumors distinguished by varying clinicopathologic and molecular features. Inter- and intra-tumor heterogeneity complicate efforts to classify, grade, and prognosticate. The Cancer Genome Atlas (TCGA) comprehensively profiled 7 sarcoma subtypes, providing detailed insight into their molecular features. However, it is not practical to implement such resource-intensive genomic methods clinically. Recent advances in deep learning techniques allow automated detection of imaging-based biomarkers. We study correlations between molecular and imaging features in their ability to classify sarcomas and predict clinical behavior.


Methods:

Hematoxylin & Eosin-stained TCGA slides were divided into tiles and fed into Inception v3 Deep Neural Network outputting 2,048 features/tile. These features were processed by a novel algorithm to calculate tile-level heterogeneity scores (fig 1). Patient-level tumor heterogeneity score (THS) was defined as the 95%-ile of tile-level scores from each patient’s pooled slides.


Results:

642 images from 206 sarcomas were divided into 1,842,044 tiles. THS positively correlated with mutation load (p=0.04) and mitotic rate (p=0.01). Grade 2/3 tumors showed significantly higher THS compared to grade 1 (p=0.02) as did tumors with vs. without necrosis (p< 0.001). THS also correlated significantly with TCGA-derived mRNA and miRNA clusters (p=0.04 and p=0.04). Finally, the score approximated TCGA-reported subtype-specific survival curves, for example showing significant separation in dedifferentiated liposarcoma outcome by THS (p=0.01).


Conclusions:

Imaging-based biomarkers are informative complements to molecular data using slides produced in routine clinical workflow without need for additional assays. Deep learning features can quantify tumor heterogeneity, creating a composite biomarker that correlates with traditional histopathological features and mirrors the performance of TCGA-reported metrics. As larger combined molecular- and imaging-datasets become available, deep learning classifiers can be trained to identify predictive biomarkers with high accuracy using images alone and may obviate the need for specialized molecular testing.