Slideflow: deep learning for digital histopathology with real-time whole-slide visualization.

Deep learning methods have emerged as powerful tools for analyzing histopathological images, but current methods are often specialized for specific domains and software environments, and few open-source options exist for deploying models in an interactive interface. Experimenting with different deep learning approaches typically requires switching software libraries and reprocessing data, reducing the feasibility and practicality of experimenting with new architectures. We developed a flexible deep learning library for histopathology called Slideflow, a package which supports a broad array of deep learning methods for digital pathology and includes a fast whole-slide interface for deploying trained models. Slideflow includes unique tools for whole-slide image data processing, efficient stain normalization and augmentation, weakly-supervised whole-slide classification, uncertainty quantification, feature generation, feature space analysis, and explainability. Whole-slide image processing is highly optimized, enabling whole-slide tile extraction at 40x magnification in 2.5 s per slide. The framework-agnostic data processing pipeline enables rapid experimentation with new methods built with either Tensorflow or PyTorch, and the graphical user interface supports real-time visualization of slides, predictions, heatmaps, and feature space characteristics on a variety of hardware devices, including ARM-based devices such as the Raspberry Pi.

Machine learning for detection and classification of oral potentially malignant disorders: A conceptual review.

Oral potentially malignant disorders represent precursor lesions that may undergo malignant transformation to oral cancer. There are many known risk factors associated with the development of oral potentially malignant disorders, and contribute to the risk of malignant transformation. Although many advances have been reported to understand the biological behavior of oral potentially malignant disorders, their clinical features that indicate the characteristics of malignant transformation are not well established. Early diagnosis of malignancy is the most important factor to improve patients' prognosis. The integration of machine learning into routine diagnosis has recently emerged as an adjunct to aid clinical examination. Increased performances of artificial intelligence AI-assisted medical devices are claimed to exceed the human capability in the clinical detection of early cancer. Therefore, the aim of this narrative review is to introduce artificial intelligence terminology, concepts, and models currently used in oncology to familiarize oral medicine scientists with the language skills, best research practices, and knowledge for developing machine learning models applied to the clinical detection of oral potentially malignant disorders.

Induction Chemotherapy in Low-Risk HPV+ Oropharyngeal Cancer.

OPINION STATEMENT: Human papillomavirus (HPV) associated oropharyngeal squamous cell carcinoma (OPSCC) is rapidly increasing in incidence, and has now become the most common head and neck cancer (HNC). Studies have demonstrated that HPV associated OPSCC is associated with a favorable prognosis compared with its HPV-negative counterparts, yet standard multimodality therapy is often associated with substantial acute and late treatment-related toxicity. While locoregional control is improved in HPV+ OPSCC, distant metastasis rate has gained recognition as a major cause of death in this population, with some studies suggesting similar rates as non-HPV-related cancers. Induction chemotherapy has been of long-standing interest in locoregionally advanced HNC, yet its use in combination with concomitant chemoradiation remains an area of controversy as a survival benefit remains unproven following randomized trials. Nevertheless, response to induction chemotherapy remains an important dynamic and prognostic biomarker, with response-adaptive de-intensified therapy in HPV+ OPSCC gaining traction in single-arm phase II studies demonstrating promising results. The emergence of immunotherapy in the recurrent/metastatic setting for HNC has led to enthusiasm to incorporate in the curative setting, yet its role remains undefined. Our institutional paradigm for HPV+ OPSCC incorporates induction therapy followed by risk and response adaptive locoregional treatment. Ultimately, the role of induction therapy in HPV+ OPSCC will need to be investigated in a randomized setting to be incorporated routinely into clinical practice.

Deep learning prediction of BRAF-RAS gene expression signature identifies noninvasive follicular thyroid neoplasms with papillary-like nuclear features.

Noninvasive follicular thyroid neoplasms with papillary-like nuclear features (NIFTP) are follicular-patterned thyroid neoplasms defined by nuclear atypia and indolent behavior. They harbor RAS mutations, rather than BRAFV600E mutations as is observed in papillary thyroid carcinomas with extensive follicular growth. Reliably identifying NIFTPs aids in safe therapy de-escalation, but has proven to be challenging due to interobserver variability and morphologic heterogeneity. The genomic scoring system BRS (BRAF-RAS score) was developed to quantify the extent to which a tumor's expression profile resembles a BRAFV600E or RAS-mutant neoplasm. We proposed that deep learning prediction of BRS could differentiate NIFTP from other follicular-patterned neoplasms. A deep learning model was trained on slides from a dataset of 115 thyroid neoplasms to predict tumor subtype (NIFTP, PTC-EFG, or classic PTC), and was used to generate predictions for 497 thyroid neoplasms within The Cancer Genome Atlas (TCGA). Within follicular-patterned neoplasms, tumors with positive BRS (RAS-like) were 8.5 times as likely to carry an NIFTP prediction than tumors with negative BRS (89.7% vs 10.5%, P < 0.0001). To test the hypothesis that BRS may serve as a surrogate for biological processes that determine tumor subtype, a separate model was trained on TCGA slides to predict BRS as a linear outcome. This model performed well in cross-validation on the training set (R2 = 0.67, dichotomized AUC = 0.94). In our internal cohort, NIFTPs were near universally predicted to have RAS-like BRS; as a sole discriminator of NIFTP status, predicted BRS performed with an AUC of 0.99 globally and 0.97 when restricted to follicular-patterned neoplasms. BRAFV600E-mutant PTC-EFG had BRAFV600E-like predicted BRS (mean -0.49), nonmutant PTC-EFG had more intermediate predicted BRS (mean -0.17), and NIFTP had RAS-like BRS (mean 0.35; P < 0.0001). In summary, histologic features associated with the BRAF-RAS gene expression spectrum are detectable by deep learning and can aid in distinguishing indolent NIFTP from PTCs.

Myc and ChREBP transcription factors cooperatively regulate normal and neoplastic hepatocyte proliferation in mice.

Analogous to the c-Myc (Myc)/Max family of bHLH-ZIP transcription factors, there exists a parallel regulatory network of structurally and functionally related proteins with Myc-like functions. Two related Myc-like paralogs, termed MondoA and MondoB/carbohydrate response element-binding protein (ChREBP), up-regulate gene expression in heterodimeric association with the bHLH-ZIP Max-like factor Mlx. Myc is necessary to support liver cancer growth, but not for normal hepatocyte proliferation. Here, we investigated ChREBP's role in these processes and its relationship to Myc. Unlike Myc loss, ChREBP loss conferred a proliferative disadvantage to normal murine hepatocytes, as did the combined loss of ChREBP and Myc. Moreover, hepatoblastomas (HBs) originating in myc-/-, chrebp-/-, or myc-/-/chrebp-/- backgrounds grew significantly more slowly. Metabolic studies on livers and HBs in all three genetic backgrounds revealed marked differences in oxidative phosphorylation, fatty acid ß-oxidation (FAO), and pyruvate dehydrogenase activity. RNA-Seq of livers and HBs suggested seven distinct mechanisms of Myc-ChREBP target gene regulation. Gene ontology analysis indicated that many transcripts deregulated in the chrebp-/- background encode enzymes functioning in glycolysis, the TCA cycle, and ß- and ω-FAO, whereas those dysregulated in the myc-/- background encode enzymes functioning in glycolysis, glutaminolysis, and sterol biosynthesis. In the myc-/-/chrebp-/- background, additional deregulated transcripts included those involved in peroxisomal ß- and α-FAO. Finally, we observed that Myc and ChREBP cooperatively up-regulated virtually all ribosomal protein genes. Our findings define the individual and cooperative proliferative, metabolic, and transcriptional roles for the "Extended Myc Network" under both normal and neoplastic conditions.

Sequential adaptive changes in a c-Myc-driven model of hepatocellular carcinoma.

Hepatocellular carcinoma (HCC) is a common cancer that frequently overexpresses the c-Myc (Myc) oncoprotein. Using a mouse model of Myc-induced HCC, we studied the metabolic, biochemical, and molecular changes accompanying HCC progression, regression, and recurrence. These involved altered rates of pyruvate and fatty acid ß-oxidation and the likely re-directing of glutamine into biosynthetic rather than energy-generating pathways. Initial tumors also showed reduced mitochondrial mass and differential contributions of electron transport chain complexes I and II to respiration. The uncoupling of complex II's electron transport function from its succinate dehydrogenase activity also suggested a mechanism by which Myc generates reactive oxygen species. RNA sequence studies revealed an orderly progression of transcriptional changes involving pathways pertinent to DNA damage repair, cell cycle progression, insulin-like growth factor signaling, innate immunity, and further metabolic re-programming. Only a subset of functions deregulated in initial tumors was similarly deregulated in recurrent tumors thereby indicating that the latter can "normalize" some behaviors to suit their needs. An interactive and freely available software tool was developed to allow continued analyses of these and other transcriptional profiles. Collectively, these studies define the metabolic, biochemical, and molecular events accompanyingHCCevolution, regression, and recurrence in the absence of any potentially confounding therapies.

Diagnostic and prognostic implications of ribosomal protein transcript expression patterns in human cancers.

BACKGROUND: Ribosomes, the organelles responsible for the translation of mRNA, are comprised of four rRNAs and ~ 80 ribosomal proteins (RPs). Although canonically assumed to be maintained in equivalent proportions, some RPs have been shown to possess differential expression across tissue types. Dysregulation of RP expression occurs in a variety of human diseases, notably in many cancers, and altered expression of some RPs correlates with different tumor phenotypes and patient survival. Little work has been done, however, to characterize overall patterns of RP transcript (RPT) expression in human cancers. METHODS: To investigate the impact of global RPT expression patterns on tumor phenotypes, we analyzed RPT expression of ~ 10,000 human tumors and over 700 normal tissues from The Cancer Genome Atlas (TCGA) using t-distributed stochastic neighbor embedding (t-SNE). Clusters of tumors identified by t-SNE were then analyzed with chi-squared and t-tests to compare phenotypic data, ANOVA to compare individual RPT expression, and Kaplan-Meier curves to assess survival differences. RESULTS: Normal tissues and cancers possess distinct and readily discernible RPT expression patterns that are independent of their absolute levels of expression. In tumors, RPT patterning is distinct from that of normal tissues, identifies heretofore unrecognized tumor subtypes, and in many cases correlates with molecular, pathological, and clinical features, including survival. CONCLUSIONS: RPT expression patterns are both tissue-specific and tumor-specific. These could be used as a powerful and novel method of tumor classification, offering a potential clinical tool for prognosis and therapeutic stratification.

Coordinated Activities of Multiple Myc-dependent and Myc-independent Biosynthetic Pathways in Hepatoblastoma.

Hepatoblastoma (HB) is associated with aberrant activation of the ß-catenin and Hippo/YAP signaling pathways. Overexpression of mutant ß-catenin and YAP in mice induces HBs that express high levels of c-Myc (Myc). In light of recent observations that Myc is unnecessary for long-term hepatocyte proliferation, we have now examined its role in HB pathogenesis using the above model. Although Myc was found to be dispensable for in vivo HB initiation, it was necessary to sustain rapid tumor growth. Gene expression profiling identified key molecular differences between myc+/+ (WT) and myc-/- (KO) hepatocytes and HBs that explain these behaviors. In HBs, these included both Myc-dependent and Myc-independent increases in families of transcripts encoding ribosomal proteins, non-structural factors affecting ribosome assembly and function, and enzymes catalyzing glycolysis and lipid bio-synthesis. In contrast, transcripts encoding enzymes involved in fatty acid ß-oxidation were mostly down-regulated. Myc-independent metabolic changes associated with HBs included dramatic reductions in mitochondrial mass and oxidative function, increases in ATP content and pyruvate dehydrogenase activity, and marked inhibition of fatty acid ß-oxidation (FAO). Myc-dependent metabolic changes included higher levels of neutral lipid and acetyl-CoA in WT tumors. The latter correlated with higher histone H3 acetylation. Collectively, our results indicate that the role of Myc in HB pathogenesis is to impose mutually dependent changes in gene expression and metabolic reprogramming that are unattainable in non-transformed cells and that cooperate to maximize tumor growth.

Generative Adversarial Networks Accurately Reconstruct Pan-Cancer Histology from Pathologic, Genomic, and Radiographic Latent Features.

Artificial intelligence models have been increasingly used in the analysis of tumor histology to perform tasks ranging from routine classification to identification of novel molecular features. These approaches distill cancer histologic images into high-level features which are used in predictions, but understanding the biologic meaning of such features remains challenging. We present and validate a custom generative adversarial network - HistoXGAN - capable of reconstructing representative histology using feature vectors produced by common feature extractors. We evaluate HistoXGAN across 29 cancer subtypes and demonstrate that reconstructed images retain information regarding tumor grade, histologic subtype, and gene expression patterns. We leverage HistoXGAN to illustrate the underlying histologic features for deep learning models for actionable mutations, identify model reliance on histologic batch effect in predictions, and demonstrate accurate reconstruction of tumor histology from radiographic imaging for a 'virtual biopsy'.

Artificial intelligence-based morphologic classification and molecular characterization of neuroblastic tumors from digital histopathology.

A deep learning model using attention-based multiple instance learning (aMIL) and self-supervised learning (SSL) was developed to perform pathologic classification of neuroblastic tumors and assess MYCN-amplification status using H&E-stained whole slide digital images. The model demonstrated strong performance in identifying diagnostic category, grade, mitosis-karyorrhexis index (MKI), and MYCN-amplification on an external test dataset. This AI-based approach establishes a valuable tool for automating diagnosis and precise classification of neuroblastoma tumors.

Evaluation of Major Pathologic Response and Pathologic Complete Response as Surrogate End Points for Survival in Randomized Controlled Trials of Neoadjuvant Immune Checkpoint Blockade in Resectable in NSCLC.

INTRODUCTION: Controversy remains as to whether pathologic complete response (pCR) and major pathologic response (MPR) represent surrogate end points for event-free survival (EFS) and overall survival (OS) in neoadjuvant trials for resectable NSCLC. METHODS: A search of PubMed and archives of international conference abstracts was performed from June 2017 through October 31, 2023. Studies incorporating a neoadjuvant arm with immune checkpoint blockade alone or in combination with chemotherapy were included. Those not providing information regarding pCR, MPR, EFS, or OS were excluded. For trial-level surrogacy, log ORs for pCR and MPR and log hazard ratios for EFS and OS were analyzed using a linear regression model weighted by sample size. The regression coefficient and R2 with 95% confidence interval were calculated by the bootstrapping approach. RESULTS: Seven randomized clinical trials were identified for a total of 2385 patients. At the patient level, the R2 of pCR and MPR with 2-year EFS were 0.82 (0.66-0.94) and 0.81 (0.63-0.93), respectively. The OR of 2-year EFS rates by response status was 0.12 (0.07-0.19) and 0.11 (0.05-0.22), respectively. For the 2-year OS, the R2 of pCR and MPR were 0.55 (0.09-0.98) and 0.52 (0.10-0.96), respectively. At the trial level, the R2 for the association of OR for response and HR for EFS was 0.58 (0.00-0.97) and 0.61 (0.00-0.97), respectively. CONCLUSIONS: Our analyses reveal a robust correlation between pCR and MPR with 2-year EFS but not OS. Trial-level surrogacy was moderate but imprecise. More mature follow-up and data to assess the impact of study crossover are needed.

Applications of Deep Learning in Endocrine Neoplasms.

Machine learning methods have been growing in prominence across all areas of medicine. In pathology, recent advances in deep learning (DL) have enabled computational analysis of histological samples, aiding in diagnosis and characterization in multiple disease areas. In cancer, and particularly endocrine cancer, DL approaches have been shown to be useful in tasks ranging from tumor grading to gene expression prediction. This review summarizes the current state of DL research in endocrine cancer histopathology with an emphasis on experimental design, significant findings, and key limitations.

Data augmentation and multimodal learning for predicting drug response in patient-derived xenografts from gene expressions and histology images.

Patient-derived xenografts (PDXs) are an appealing platform for preclinical drug studies. A primary challenge in modeling drug response prediction (DRP) with PDXs and neural networks (NNs) is the limited number of drug response samples. We investigate multimodal neural network (MM-Net) and data augmentation for DRP in PDXs. The MM-Net learns to predict response using drug descriptors, gene expressions (GE), and histology whole-slide images (WSIs). We explore whether combining WSIs with GE improves predictions as compared with models that use GE alone. We propose two data augmentation methods which allow us training multimodal and unimodal NNs without changing architectures with a single larger dataset: 1) combine single-drug and drug-pair treatments by homogenizing drug representations, and 2) augment drug-pairs which doubles the sample size of all drug-pair samples. Unimodal NNs which use GE are compared to assess the contribution of data augmentation. The NN that uses the original and the augmented drug-pair treatments as well as single-drug treatments outperforms NNs that ignore either the augmented drug-pairs or the single-drug treatments. In assessing the multimodal learning based on the MCC metric, MM-Net outperforms all the baselines. Our results show that data augmentation and integration of histology images with GE can improve prediction performance of drug response in PDXs.

Integration of clinical features and deep learning on pathology for the prediction of breast cancer recurrence assays and risk of recurrence.

Gene expression-based recurrence assays are strongly recommended to guide the use of chemotherapy in hormone receptor-positive, HER2-negative breast cancer, but such testing is expensive, can contribute to delays in care, and may not be available in low-resource settings. Here, we describe the training and independent validation of a deep learning model that predicts recurrence assay result and risk of recurrence using both digital histology and clinical risk factors. We demonstrate that this approach outperforms an established clinical nomogram (area under the receiver operating characteristic curve of 0.83 versus 0.76 in an external validation cohort, p = 0.0005) and can identify a subset of patients with excellent prognoses who may not need further genomic testing.

What is the ideal endpoint in early-stage immunotherapy neoadjuvant trials in lung cancer?

Numerous clinical trials investigating neoadjuvant immune checkpoint inhibitors (ICI) have been performed over the last 5 years. As the number of neoadjuvant trials increases, attention must be paid to identifying informative trial endpoints. Complete pathologic response has been shown to be an appropriate surrogate endpoint for clinical outcomes, such as event-free survival or overall survival, in breast cancer and bladder cancer, but it is less established for non-small-cell lung cancer (NSCLC). The simultaneous advances reported with adjuvant ICI make the optimal strategy for early-stage disease debatable. Considering the long time required to conduct trials, it is important to identify optimal endpoints and discover surrogate endpoints for survival that can help guide ongoing clinical research. Endpoints can be grouped into two categories: medical and surgical. Medical endpoints are measures of survival and drug activity; surgical endpoints describe the feasibility of neoadjuvant approaches at a surgical level as well as perioperative attrition and complications. There are also several exploratory endpoints, including circulating tumor DNA clearance and radiomics. In this review, we outline the advantages and disadvantages of commonly reported endpoints for clinical trials of neoadjuvant regimens in NSCLC.

Deep learning generates synthetic cancer histology for explainability and education.

Artificial intelligence methods including deep neural networks (DNN) can provide rapid molecular classification of tumors from routine histology with accuracy that matches or exceeds human pathologists. Discerning how neural networks make their predictions remains a significant challenge, but explainability tools help provide insights into what models have learned when corresponding histologic features are poorly defined. Here, we present a method for improving explainability of DNN models using synthetic histology generated by a conditional generative adversarial network (cGAN). We show that cGANs generate high-quality synthetic histology images that can be leveraged for explaining DNN models trained to classify molecularly-subtyped tumors, exposing histologic features associated with molecular state. Fine-tuning synthetic histology through class and layer blending illustrates nuanced morphologic differences between tumor subtypes. Finally, we demonstrate the use of synthetic histology for augmenting pathologist-in-training education, showing that these intuitive visualizations can reinforce and improve understanding of histologic manifestations of tumor biology.

The EU-funded I3LUNG Project: Integrative Science, Intelligent Data Platform for Individualized LUNG Cancer Care With Immunotherapy.

Although immunotherapy (IO) has changed the paradigm for the treatment of patients with advanced non-small cell lung cancers (aNSCLC), only around 30% to 50% of treated patients experience a long-term benefit from IO. Furthermore, the identification of the 30 to 50% of patients who respond remains a major challenge, as programmed Death-Ligand 1 (PD-L1) is currently the only biomarker used to predict the outcome of IO in NSCLC patients despite its limited efficacy. Considering the dynamic complexity of the immune system-tumor microenvironment (TME) and its interaction with the host's and patient's behavior, it is unlikely that a single biomarker will accurately predict a patient's outcomes. In this scenario, Artificial Intelligence (AI) and Machine Learning (ML) are becoming essential to the development of powerful decision-making tools that are able to deal with this high-complexity and provide individualized predictions to better match treatments to individual patients and thus improve patient outcomes and reduce the economic burden of aNSCLC on healthcare systems. I3LUNG is an international, multicenter, retrospective and prospective, observational study of patients with aNSCLC treated with IO, entirely funded by European Union (EU) under the Horizon 2020 (H2020) program. Using AI-based tools, the aim of this study is to promote individualized treatment in aNSCLC, with the goals of improving survival and quality of life, minimizing or preventing undue toxicity and promoting efficient resource allocation. The final objective of the project is the construction of a novel, integrated, AI-assisted data storage and elaboration platform to guide IO administration in aNSCLC, ensuring easy access and cost-effective use by healthcare providers and patients.

PEOPLE (NTC03447678), a phase II trial to test pembrolizumab as first-line treatment in patients with advanced NSCLC with PD-L1 <50%: a multiomics analysis.

BACKGROUND: Chemoimmunotherapy represents the standard of care for patients with advanced non-small cell lung cancer (NSCLC) and programmed death-ligand 1 (PD-L1) <50%. Although single-agent pembrolizumab has also demonstrated some activity in this setting, no reliable biomarkers yet exist for selecting patients likely to respond to single-agent immunotherapy. The main purpose of the study was to identify potential new biomarkers associated with progression-free-survival (PFS) within a multiomics analysis. METHODS: PEOPLE (NTC03447678) was a prospective phase II trial evaluating first-line pembrolizumab in patients with advanced EGFR and ALK wild type treatment-naïve NSCLC with PD-L1 <50%. Circulating immune profiling was performed by determination of absolute cell counts with multiparametric flow cytometry on freshly isolated whole blood samples at baseline and at first radiological evaluation. Gene expression profiling was performed using nCounter PanCancer IO 360 Panel (NanoString) on baseline tissue. Gut bacterial taxonomic abundance was obtained by shotgun metagenomic sequencing of stool samples at baseline. Omics data were analyzed with sequential univariate Cox proportional hazards regression predicting PFS, with Benjamini-Hochberg multiple comparisons correction. Biological features significant with univariate analysis were analyzed with multivariate least absolute shrinkage and selection operator (LASSO). RESULTS: From May 2018 to October 2020, 65 patients were enrolled. Median follow-up and PFS were 26.4 and 2.9 months, respectively. LASSO integration analysis, with an optimal lambda of 0.28, showed that peripheral blood natural killer cells/CD56dimCD16+ (HR 0.56, 0.41-0.76, p=0.006) abundance at baseline and non-classical CD14dimCD16+monocytes (HR 0.52, 0.36-0.75, p=0.004), eosinophils (CD15+CD16-) (HR 0.62, 0.44-0.89, p=0.03) and lymphocytes (HR 0.32, 0.19-0.56, p=0.001) after first radiologic evaluation correlated with favorable PFS as well as high baseline expression levels of CD244 (HR 0.74, 0.62-0.87, p=0.05) protein tyrosine phosphatase receptor type C (HR 0.55, 0.38-0.81, p=0.098) and killer cell lectin like receptor B1 (HR 0.76, 0.66-0.89, p=0.05). Interferon-responsive factor 9 and cartilage oligomeric matrix protein genes correlated with unfavorable PFS (HR 3.03, 1.52-6.02, p 0.08 and HR 1.22, 1.08-1.37, p=0.06, corrected). No microbiome features were selected. CONCLUSIONS: This multiomics approach was able to identify immune cell subsets and expression levels of genes associated to PFS in patients with PD-L1 <50% NSCLC treated with first-line pembrolizumab. These preliminary data will be confirmed in the larger multicentric international I3LUNG trial (NCT05537922). TRIAL REGISTRATION NUMBER: 2017-002841-31.

Uncertainty-informed deep learning models enable high-confidence predictions for digital histopathology.

A model's ability to express its own predictive uncertainty is an essential attribute for maintaining clinical user confidence as computational biomarkers are deployed into real-world medical settings. In the domain of cancer digital histopathology, we describe a clinically-oriented approach to uncertainty quantification for whole-slide images, estimating uncertainty using dropout and calculating thresholds on training data to establish cutoffs for low- and high-confidence predictions. We train models to identify lung adenocarcinoma vs. squamous cell carcinoma and show that high-confidence predictions outperform predictions without uncertainty, in both cross-validation and testing on two large external datasets spanning multiple institutions. Our testing strategy closely approximates real-world application, with predictions generated on unsupervised, unannotated slides using predetermined thresholds. Furthermore, we show that uncertainty thresholding remains reliable in the setting of domain shift, with accurate high-confidence predictions of adenocarcinoma vs. squamous cell carcinoma for out-of-distribution, non-lung cancer cohorts.

The impact of site-specific digital histology signatures on deep learning model accuracy and bias.

The Cancer Genome Atlas (TCGA) is one of the largest biorepositories of digital histology. Deep learning (DL) models have been trained on TCGA to predict numerous features directly from histology, including survival, gene expression patterns, and driver mutations. However, we demonstrate that these features vary substantially across tissue submitting sites in TCGA for over 3,000 patients with six cancer subtypes. Additionally, we show that histologic image differences between submitting sites can easily be identified with DL. Site detection remains possible despite commonly used color normalization and augmentation methods, and we quantify the image characteristics constituting this site-specific digital histology signature. We demonstrate that these site-specific signatures lead to biased accuracy for prediction of features including survival, genomic mutations, and tumor stage. Furthermore, ethnicity can also be inferred from site-specific signatures, which must be accounted for to ensure equitable application of DL. These site-specific signatures can lead to overoptimistic estimates of model performance, and we propose a quadratic programming method that abrogates this bias by ensuring models are not trained and validated on samples from the same site.