Prediction of MET Overexpression in Lung Adenocarcinoma from Hematoxylin and Eosin Images.

Mesenchymal epithelial transition (MET) protein overexpression is a targetable event in non-small cell lung cancer and is the subject of active drug development. Challenges in identifying patients for these therapies include lack of access to validated testing, such as standardized immunohistochemistry assessment, and consumption of valuable tissue for a single gene/protein assay. Development of prescreening algorithms using routinely available digitized hematoxylin and eosin (H&E)-stained slides to predict MET overexpression could promote testing for those who will benefit most. Recent literature reports a positive correlation between MET protein overexpression and RNA expression. In this work, a large database of matched H&E slides and RNA expression data were leveraged to train a weakly supervised model to predict MET RNA overexpression directly from H&E images. This model was evaluated on an independent holdout test set of 300 overexpressed and 289 normal patients, demonstrating a receiver operating characteristic area under curve of 0.70 (95th percentile interval: 0.66 to 0.74) with stable performance characteristics across different patient clinical variables and robust to synthetic noise on the test set. These results suggest that H&E-based predictive models could be useful to prioritize patients for confirmatory testing of MET protein or MET gene expression status.

Replicating and identifying large cell neuroblastoma using high-dose intra-tumoral chemotherapy and automated digital analysis.

PURPOSE: Large cell neuroblastomas (LCN) are frequently seen in recurrent, high-risk neuroblastoma but are rare in primary tumors. LCN, characterized by large nuclei with prominent nucleoli, predict a poor prognosis. We hypothesize that LCN can be created with high-dose intra-tumoral chemotherapy and identified by a digital analysis system. METHODS: Orthotopic mouse xenografts were created using human neuroblastoma and treated with high-dose chemotherapy delivered locally via sustained-release silk platforms, inducing tumor remission. After recurrence, LCN populations were identified on H&E sections manually. Clusters of typical LCN and non-LCN cells were divided equally into training and test sets for digital analysis. Marker-controlled watershed segmentation was used to identify nuclei and characterize their features. Logistic regression was developed to distinguish LCN from non-LCN. RESULTS: Image analysis identified 15,000 nuclei and characterized 70 nuclear features. A 19-feature model provided AUC >0.90 and 100% accuracy when >30% nuclei/cluster were predicted as LCN. Overall accuracy was 87%. CONCLUSIONS: We recreated LCN using high-dose chemotherapy and developed an automated method for defining LCN histologically. Features in the model provide insight into LCN nuclear phenotypic changes that may be related to increased activity. This model could be adapted to identify LCN in human tumors and correlated with clinical outcomes.

Multi-Field-of-View Deep Learning Model Predicts Nonsmall Cell Lung Cancer Programmed Death-Ligand 1 Status from Whole-Slide Hematoxylin and Eosin Images.

BACKGROUND: Tumor programmed death-ligand 1 (PD-L1) status is useful in determining which patients may benefit from programmed death-1 (PD-1)/PD-L1 inhibitors. However, little is known about the association between PD-L1 status and tumor histopathological patterns. Using deep learning, we predicted PD-L1 status from hematoxylin and eosin (H and E) whole-slide images (WSIs) of nonsmall cell lung cancer (NSCLC) tumor samples. MATERIALS AND METHODS: One hundred and thirty NSCLC patients were randomly assigned to training (n = 48) or test (n = 82) cohorts. A pair of H and E and PD-L1-immunostained WSIs was obtained for each patient. A pathologist annotated PD-L1 positive and negative tumor regions on the training samples using immunostained WSIs for reference. From the H and E WSIs, over 145,000 training tiles were generated and used to train a multi-field-of-view deep learning model with a residual neural network backbone. RESULTS: The trained model accurately predicted tumor PD-L1 status on the held-out test cohort of H and E WSIs, which was balanced for PD-L1 status (area under the receiver operating characteristic curve [AUC] =0.80, P << 0.01). The model remained effective over a range of PD-L1 cutoff thresholds (AUC = 0.67-0.81, P ≤ 0.01) and when different proportions of the labels were randomly shuffled to simulate interpathologist disagreement (AUC = 0.63-0.77, P ≤ 0.03). CONCLUSIONS: A robust deep learning model was developed to predict tumor PD-L1 status from H and E WSIs in NSCLC. These results suggest that PD-L1 expression is correlated with the morphological features of the tumor microenvironment.

Integrating RNA expression and visual features for immune infiltrate prediction.

Patient responses to cancer immunotherapy are shaped by their unique genomic landscape and tumor microenvironment. Clinical advances in immunotherapy are changing the treatment landscape by enhancing a patient's immune response to eliminate cancer cells. While this provides potentially beneficial treatment options for many patients, only a minority of these patients respond to immunotherapy. In this work, we examined RNA-seq data and digital pathology images from individual patient tumors to more accurately characterize the tumor-immune microenvironment. Several studies implicate an inflamed microenvironment and increased percentage of tumor infiltrating immune cells with better response to specific immunotherapies in certain cancer types. We developed NEXT (Neural-based models for integrating gene EXpression and visual Texture features) to more accurately model immune infiltration in solid tumors. To demonstrate the utility of the NEXT framework, we predicted immune infiltrates across four different cancer types and evaluated our predictions against expert pathology review. Our analyses demonstrate that integration of imaging features improves prediction of the immune infiltrate. Of note, this effect was preferentially observed for B cells and CD8 T cells. In sum, our work effectively integrates both RNA-seq and imaging data in a clinical setting and provides a more reliable and accurate prediction of the immune composition in individual patient tumors.

Computer vision detects subtle histological effects of dutasteride on benign prostate.

OBJECTIVE: To determine whether a computer vision-based approach applied to haematoxylin and eosin (H&E) prostate biopsy images can distinguish dutasteride-treated tissue from placebo, and identify features associated with degree of responsiveness to 5α-reductase inhibitor (5ARI) therapy. SUBJECTS AND METHODS: Our study population comprised 100 treatment-adherent men without prostate cancer assigned to dutasteride or placebo in the REDUCE trial, who had slides available from mandatory year-4 biopsies. Half of the men also provided slides from a year-2 biopsy. We obtained 20× whole-slide images and used specialized software to generate a library of 1 300 epithelial and stromal features from objects comprising superpixels and several types of nuclei, including spatial relations among objects between and within each hierarchical level. We used penalized logistic regression and fivefold cross-validation to find optimal combinations of histological features in the year-4 biopsies. Feature data from the year-2 biopsies were fitted to a final model for independent validation. Two pathologists, blinded to treatment, scored each image for focal atrophy and histological features previously linked to 5AR1 treatment. RESULTS: Consensus classification by pathologists obtained a discrimination accuracy equivalent to chance. A 21-feature computer vision model gave a cross-validation area under the curve of 0.97 (95% confidence interval [CI] 0.95-0.99) in the year-4 biopsies and 0.79 (95% CI: 0.65-0.92) in the set-aside year-2 biopsies. Histology scores were not correlated with change in prostate-specific antigen level, serum dihydrotestosterone level or gland volume. Key features associated with dutasteride treatment included greater shape and colour uniformity in stroma, irregular clustering of epithelial nuclei, and greater variation in lumen shape. CONCLUSION: The present findings show that a computer vision approach can detect subtle histological effects attributable to dutasteride, resulting in a continuous measure of responsiveness to the drug that could eventually be used to predict individual patient response in the context of BPH treatment or cancer chemoprevention.

Empirical comparison of color normalization methods for epithelial-stromal classification in H and E images.

CONTEXT: Color normalization techniques for histology have not been empirically tested for their utility for computational pathology pipelines. AIMS: We compared two contemporary techniques for achieving a common intermediate goal - epithelial-stromal classification. SETTINGS AND DESIGN: Expert-annotated regions of epithelium and stroma were treated as ground truth for comparing classifiers on original and color-normalized images. MATERIALS AND METHODS: Epithelial and stromal regions were annotated on thirty diverse-appearing H and E stained prostate cancer tissue microarray cores. Corresponding sets of thirty images each were generated using the two color normalization techniques. Color metrics were compared for original and color-normalized images. Separate epithelial-stromal classifiers were trained and compared on test images. Main analyses were conducted using a multiresolution segmentation (MRS) approach; comparative analyses using two other classification approaches (convolutional neural network [CNN], Wndchrm) were also performed. STATISTICAL ANALYSIS: For the main MRS method, which relied on classification of super-pixels, the number of variables used was reduced using backward elimination without compromising accuracy, and test - area under the curves (AUCs) were compared for original and normalized images. For CNN and Wndchrm, pixel classification test-AUCs were compared. RESULTS: Khan method reduced color saturation while Vahadane reduced hue variance. Super-pixel-level test-AUC for MRS was 0.010-0.025 (95% confidence interval limits ± 0.004) higher for the two normalized image sets compared to the original in the 10-80 variable range. Improvement in pixel classification accuracy was also observed for CNN and Wndchrm for color-normalized images. CONCLUSIONS: Color normalization can give a small incremental benefit when a super-pixel-based classification method is used with features that perform implicit color normalization while the gain is higher for patch-based classification methods for classifying epithelium versus stroma.