The level of PD-L1 expression in immunohistochemistry (IHC) assays is a key biomarker for the identification of Non-Small-Cell-Lung-Cancer (NSCLC) patients that may respond to anti PD-1/PD-L1 treatments. The quantification of PD-L1 expression currently includes the visual estimation by a pathologist of the percentage (tumor proportional scoring or TPS) of tumor cells showing PD-L1 staining. Known challenges like differences in positivity estimation around clinically relevant cut-offs and sub-optimal quality of samples makes visual scoring tedious and subjective, yielding a scoring variability between pathologists. In this work, we propose a novel deep learning solution that enables the first automated and objective scoring of PD-L1 expression in late stage NSCLC needle biopsies. To account for the low amount of tissue available in biopsy images and to restrict the amount of manual annotations necessary for training, we explore the use of semi-supervised approaches against standard fully supervised methods. We consolidate the manual annotations used for training as well the visual TPS scores used for quantitative evaluation with multiple pathologists. Concordance measures computed on a set of slides unseen during training provide evidence that our automatic scoring method matches visual scoring on the considered dataset while ensuring repeatability and objectivity.

We report the ability of two deep learningbased decision systems to stratify non-small cell lung cancer (NSCLC) patients treated with checkpoint inhibitor therapy into two distinct survival groups. Both systems analyze functional and morphological properties of epithelial regions in digital histopathology whole slide images stained with the SP263 PD-L1 antibody. The first system learns to replicate the pathologist assessment of the Tumor Cell (TC) score with a cut-point for positivity at 25% for patient stratification. The second system is free from assumptions related to TC scoring and directly learns patient stratification from the overall survival time and event information. Both systems are built on a novel unpaired domain adaptation deep learning solution for epithelial region segmentation. This approach significantly reduces the need for large pixel-precise manually annotated datasets while superseding serial sectioning or re-staining of slides to obtain ground truth by cytokeratin staining. The capacity of the first system to replicate the TC scoring by pathologists is evaluated on 703 unseen cases, with an addition of 97 cases from an independent cohort. Our results show Lin's concordance values of 0.93 and 0.96 against pathologist scoring, respectively. The ability of the first and second system to stratify anti-PD-L1 treated patients is evaluated on 151 clinical samples. Both systems show similar stratification powers (first system: HR =0.539, p=0.004 and second system: HR=0.525, p=0.003) compared to TC scoring by pathologists (HR=0.574, p=0.01).

Background T-DXd (Enhertu®) is an FDA-approved antibody-drug conjugate (ADC) targeting HER2. T-DXd has shown anti-tumor activity, not only in patients with HER2-overexpressing (IHC3+/2+ ISH+) breast cancer (BC) but also in patients with BC with low HER2 expression (IHC1+/2+ ISH−). Current HER2 protein expression assessment is based on manual pathologist scoring that classifies tumors by the percentage of tumor cells with highest intensity and completeness of staining. A critical need exists for more objective and quantitative methods to assess HER2 expression, specifically to better identify patients with low-level expression if T-DXd proves to be efficacious in this patient population. Methods We used deep learning (DL)-based image analysis (IA) to generate a novel HER2 Quantitative Continuous Score (QCS). Data analytic techniques determined optimal HER2 QCS for the J101 trial (NCT02564900) of 151 patients with varying HER2 expression levels (1+, 2+, 3+). HER2 QCS consists of DL models to detect membrane, cytoplasm, and nuclei of all tumor cells. QCS was extensively trained using pathologists’ annotations, and the performance was validated on unseen data to ensure its generalization and robustness. QCS was blindly applied to J101 data. The optical density (OD; level of brown stain intensity) was computed on detected membrane to derive features that could be linked to survival prediction. QCS features were selected to maximize ORR in positive group, minimize ORR in negative group maintaining while high prevalence in the positive group. Results Analytical validation showed high correlation between QCS from automatically detected membranes and QCS from those annotated by pathologists (R=0.993). This is in the same range as correlation between three pathologists (R=0.995). HER2 QCS was largely consistent with pathologist HER2 scoring as well but showed broad quantitative overlap between IHC and ISH categories. HER2 QCS showed a direct linear relationship between ORR and increased HER2 expression across the entire assay range. In the HER2-low population (n = 65), for whom HER2-targeting therapies are not currently approved, 42% of patients responded to T-DXd, with a median PFS (mPFS) of 11 mo. Using HER2 QCS, we were able to further stratify this population into a subgroup of QCS-high patients (above a staining intensity cut-off determined by IA), with response and mPFS increased to 53% (95% CI: 36%-68%) and 14.5 mo (95% CI: 10.9 mo-NR) respectively, while the QCS-low group only showed ORR of 24% (95% CI: 9%-45%) and mPFS of 8.6 mo (95% CI: 4.2 mo-NR). Generally, best-performing QCS cutoffs were driven by most tumor cells expressing a minimal amount of HER2, in contrast to current clinical guidelines that are driven by a minority of cells expressing higher levels of HER2. We also examined spatial heterogeneity by characterizing cells as either bearing membrane stain above a determined OD threshold (positive cell) or lying within certain distances from a positive cell. We observed similar efficacy with best performing-cutoffs, again, being found when a minimal level of HER2 expression (OD) was examined. Conclusions Taken together, these data establish a first proof-of-concept demonstrating that use of HER2 QCS can potentially enhance prediction of patient outcome with T-DXd by increasing sensitivity and specificity of response, especially in the HER2-low population. The ability to identify patients in the HER2-low group who could benefit from T-DXd is critical for its use in a patient population with a high unmet need that would otherwise not be treated with anti-HER2 therapy. Further clinical verification and validation is ongoing. Citation Format: Mark Gustavson, Susanne Haneder, Andreas Spitzmueller, Ansh Kapil, Katrin Schneider, Fabiola Cecchi, Sriram Sridhar, Guenter Schmidt, Sotirios Lakis, Regina Teichert, Anatoliy Shumilov, Ana Hidalgo-Sastre, Magdalena Wienken, Hadassah Sade, J. Carl Barrett, Danielle Carroll. Novel approach to HER2 quantification: Digital pathology coupled with AI-based image and data analysis delivers objective and quantitative HER2 expression analysis for enrichment of responders to trastuzumab deruxtecan (T-DXd; DS-8201), specifically in HER2-low patients [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PD6-01.

BackgroundThe pathologist’s visual assessment of tumor proportion score (TPS) with 25% cutoff on PD-L1 stained tissue samples is an established method to select metastatic NSCLC patients that are likely to respond to an anti-PD-L1 monotherapy.1 However, manual scoring is often subject to subjectivity in human perception2 and there remains a critical need for more objective and quantitative methods to assess PD-L1 expression in immuno-oncology.MethodsWe used deep learning (DL) based image analysis (IA) to generate a novel PD-L1 Quantitative Continuous Score (QCS)3 in tumor cells. PD-L1 QCS consists of two DL models to first segment epithelial regions and second detect membranes, cytoplasm and nuclei of each tumor cell in PD-L1 immunohistochemically (IHC) stained tissue slides. The PD-L1 expression of each tumor cell compartment was estimated by the respective optical density (OD) of DAB, and tumor cells with a membrane OD greater than ODmin are considered as PD-L1-positive. A slide comprising at greater percentage of PD-L1-positive tumor cells than a cutoff value (CoV) is considered QCS-positive. The ODmin and CoV parameters were linked to patient overall survival (OS), by minimizing the Kaplan Meier log-rank p-values and keeping at least 50% prevalence in the QCS-positive subgroup.Fully supervised QCS-IA models were extensively trained using pathologists’ annotations and the performance was validated on unseen data to ensure its generalization and robustness.3 4 The QCS IA was locked and blindly applied on clinical trial data (NCT01693562, durvalumab-treated late-stage NSCLC cohort) without further refinement.ResultsData analytics techniques were used to determine optimal PD-L1 QCS parameters for the clinical trial cohort of N=162 late-stage NSCLC patients. A PD-L1 QCS algorithm (ODmin=8, CoV=57%) is able to stratify durvalumab-treated NSCLC patients at a higher prevalence and more significant log rank p-value (64%, p=0.0001) for OS (figure 1) compared to pathologist TPS (59%, p=0.01). Median OS times of (19.2 months vs 7.9 months) was observed in the QCS-positive vs negative subgroups, respectively. The box plots (figure 2) indicate an overall good agreement (72% concordance) of the fully automated QCS with the pathologists TPS, which quantitatively supports the positive visual assessment of the cell segmentation accuracy.Abstract 365 Figure 1Kaplan Meier (KM) curves for OS stratification. KM curves for Overall Survival (OS) stratification with (left) manual PD-L1 TPS score (25% cutoff), and (right) automated QCS (57% cutoff).Abstract 365 Figure 2QCS scores within TPS positive and negative groups. Box plot indicating percent positive cells (OD≥8) as measured by PD-L1 QCS within the PD-L1 high (red) and low (blue) groups as per pathologist assessment by TPS.ConclusionsThe novel Quantitative Continuous Scoring (QCS) provides an objective way of correlating a quantitative estimate of PD-L1 IHC expression on tumor cells with survival of durvalumab-treated NSCLC patients. This data establishes a first proof-of-concept demonstrating the potential utility of PD-L1 QCS towards precision medicine in immuno-oncology.ReferencesRebelatto M, et al. Development of a programmed cell death ligand-1 immunohistochemical assay validated for analysis of non-small cell lung cancer and head and neck squamous cell carcinoma. Diagnostic Pathology 2016.Tsao M S, et al. PD-L1 immunohistochemistry comparability study in real-life clinical samples: results of blueprint phase 2 project. Journal of Thoracic Oncology 2018.Gustavson M, et al. Novel approach to HER2 quantification: digital pathology coupled with AI-based image and data analysis delivers objective and quantitative HER2 expression analysis for enrichment of responders to trastuzumab deruxtecan (T-DXd; DS-8201), specifically in HER2-low patients. (2021) DOI: 10.1158/1538-7445.SABCS20-PD6-01Kapil A, et al. Domain adaptation-based deep learning for automated tumor cell (TC) scoring and survival analysis on PD-L1 stained tissue images. IEEE Transactions on Medical Imaging DOI: 10.1109/TMI.2021.3081396Ethics ApprovalClinical study NCT01693562, from which data in this report were obtained, was carried out in accordance with the Declaration of Helsinki and Good Clinical Practice guidelines. The study protocol, amendments, and participant informed consent document were approved by the appropriate institutional review boards.

Webly-supervised learning has recently emerged as an alternative paradigm to traditional supervised learning based on large-scale datasets with manual annotations.The key idea is that models such as CNNs can be learned from the noisy visual data available on the web. In this work we aim to exploit web data for video understanding tasks such as action recognition and detection. One of the main problems in webly-supervised learning is cleaning the noisy labeled data from the web. The state-of-theart paradigm relies on training a first classifier on noisy data that is then used to clean the remaining dataset. Our key insight is that this procedure biases the second classifier towards samples that the first one understands. Here we train two independent CNNs, a RGB network on web images and video frames and a second network using temporal information from optical flow. We show that training the networks independently is vastly superior to selecting the frames for the flow classifier by using our RGB network. Moreover, we show benefits in enriching the training set with different data sources from heterogeneous public web databases. We demonstrate that our framework outperforms all other webly-supervised methods on two public benchmarks, UCF-101 and Thumos'14.