Domain Adaptation-Based Deep Learning for Automated Tumor Cell (TC) Scoring and Survival Analysis on PD-L1 Stained Tissue Images.

We report the ability of two deep learning-based 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).

Comparison of continuous measures across diagnostic PD-L1 assays in non-small cell lung cancer using automated image analysis.

Tumor programmed cell death ligand-1 (PD-L1) expression is a key biomarker to identify patients with non-small cell lung cancer who may have an enhanced response to anti-programmed cell death-1 (PD-1)/PD-L1 treatment. Such treatments are used in conjunction with PD-L1 diagnostic immunohistochemistry assays. We developed a computer-aided automated image analysis with customized PD-L1 scoring algorithm that was evaluated via correlation with manual pathologist scores and used to determine comparability across PD-L1 immunohistochemistry assays. The image analysis scoring algorithm was developed to quantify the percentage of PD-L1 positive tumor cells on scans of whole-slide images of archival tumor samples from commercially available non-small cell lung cancer cases, stained with four immunohistochemistry PD-L1 assays (Ventana SP263 and SP142 and Dako 22C3 and 28-8). The scans were co-registered and tumor and exclusion annotations aligned to ensure that analysis of each case was restricted to comparable tissue areas. Reference pathologist scores were available from previous studies. F1, a statistical measure of precision and recall, and overall percentage agreement scores were used to assess concordance between pathologist and image analysis scores and between immunohistochemistry assays. In total, 471 PD-L1-evalulable samples were amenable to image analysis scoring. Image analysis and pathologist scores were highly concordant, with F1 scores ranging from 0.8 to 0.9 across varying matched PD-L1 cutoffs. Based on F1 and overall percentage agreement scores (both manual and image analysis scoring), the Ventana SP263 and Dako 28-8 and 22C3 assays were concordant across a broad range of cutoffs; however, the Ventana SP142 assay showed very different characteristics. In summary, a novel automated image analysis scoring algorithm was developed that was highly correlated with pathologist scores. The algorithm permitted quantitative comparison of existing PD-L1 diagnostic assays, confirming previous findings that indicate a high concordance between the Ventana SP263 and Dako 22C3 and 28-8 PD-L1 immunohistochemistry assays.

Interobserver Reliability of Programmed Cell Death Ligand-1 Scoring Using the VENTANA PD-L1 (SP263) Assay in NSCLC.

INTRODUCTION: The VENTANA PD-L1 (SP263) Assay is approved for use with anti-programmed cell death-1/programmed cell death ligand-1 (PD-1/PD-L1) therapies in NSCLC and urothelial carcinoma. Here, we investigate interobserver reliability of the SP263 assay, applied to PD-L1 scoring of tumor cells (TCs) in NSCLC. METHODS: Six practicing European pulmonary pathologists independently scored the proportion of TCs expressing PD-L1 (TC score) from 200 archival, commercially sourced, formalin-fixed paraffin-embedded NSCLC resections stained using the SP263 assay. Agreement in scores was analyzed using the intraclass correlation coefficient and concordance in patient's classification using Fleiss' kappa. RESULTS: Results from 172 samples showed strong pair-wise correlations between pathologists (R2 >0.89) for TC scoring with an intraclass correlation coefficient of 0.96. Overall agreement was greater than 90% for TC of 1% and above, and greater than 94% for TCs of at least 25% and at least 50%. Fleiss' kappa showed substantial agreement for TC of 1% and above, and almost perfect agreement for TCs of at least 25% and at least 50%. CONCLUSIONS: Assessment of TC score in NSCLC was highly reproducible using the SP263 assay, building confidence in the accuracy of this assay in selection of patients for anti-PD-1/PD-L1 therapy.

Relevance of deep learning to facilitate the diagnosis of HER2 status in breast cancer.

Tissue biomarker scoring by pathologists is central to defining the appropriate therapy for patients with cancer. Yet, inter-pathologist variability in the interpretation of ambiguous cases can affect diagnostic accuracy. Modern artificial intelligence methods such as deep learning have the potential to supplement pathologist expertise to ensure constant diagnostic accuracy. We developed a computational approach based on deep learning that automatically scores HER2, a biomarker that defines patient eligibility for anti-HER2 targeted therapies in breast cancer. In a cohort of 71 breast tumour resection samples, automated scoring showed a concordance of 83% with a pathologist. The twelve discordant cases were then independently reviewed, leading to a modification of diagnosis from initial pathologist assessment for eight cases. Diagnostic discordance was found to be largely caused by perceptual differences in assessing HER2 expression due to high HER2 staining heterogeneity. This study provides evidence that deep learning aided diagnosis can facilitate clinical decision making in breast cancer by identifying cases at high risk of misdiagnosis.

Gadofosveset-based biomarker of tissue albumin concentration: Technical validation in vitro and feasibility in vivo.

PURPOSE: There is currently no adequate method of mapping physiologic and pathophysiologic tissue albumin concentrations in human subjects. The objective of this study was to devise and evaluate a biomarker of regional albumin concentration using gadofosveset-enhanced MRI. THEORY AND METHODS: A binding and relaxation model was devised and evaluated in vitro in solutions of albumin at 3.0 Tesla (T) and 4.7T. The method was evaluated in the heart in seven volunteers at 3.0T. RESULTS: MRI-derived estimates of albumin concentration were in good agreement with true values over the range 0.1-1.0 mM (Pearson correlation coefficients of 0.85 and 0.88 for 3.0T and 4.7T, respectively). The mean calculated albumin concentration in the myocardium for the volunteers was 0.02 mM (range, 0.01-0.03 mM). CONCLUSION: Accurate estimates of albumin concentration in vitro suggest this may be a viable noninvasive alternative to existing techniques. In the myocardium the MRI-derived estimates of albumin concentration indicate the practical feasibility of the technique but were below expected values. Gadofosveset-enhanced MR relaxometry has potential in providing biomarkers of regional albumin concentration; further evaluation is required before it can be used reliably in vivo.

Noninvasive tumor hypoxia measurement using magnetic resonance imaging in murine U87 glioma xenografts and in patients with glioblastoma.

PURPOSE: There is a clinical need for noninvasive, nonionizing imaging biomarkers of tumor hypoxia and oxygenation. We evaluated the relationship of T1 -weighted oxygen-enhanced magnetic resonance imaging (OE-MRI) measurements to histopathology measurements of tumor hypoxia in a murine glioma xenograft and demonstrated technique translation in human glioblastoma multiforme. METHODS: Preclinical evaluation was performed in a subcutaneous murine human glioma xenograft (U87MG). Animals underwent OE-MRI followed by dynamic contrast-enhanced MRI (DCE-MRI) and histological measurement including reduced pimonidazole adducts and CD31 staining. Area under the curve (AUC) was measured for the R1 curve for OE-MRI and the gadolinium concentration curve for DCE-MRI. Clinical evaluation in five patients used analogous imaging protocols and analyses. RESULTS: Changes in AUC of OE-MRI (AUCOE ) signal were regionally heterogeneous across all U87MG tumors. Tumor regions with negative AUCOE typically had low DCE-MRI perfusion, had positive correlation with hypoxic area (P = 0.029), and had negative correlation with vessel density (P = 0.004). DCE-MRI measurements did not relate to either hypoxia or vessel density in U87MG tumors. Clinical data confirmed comparable signal changes in patients with glioblastoma. CONCLUSION: These data support further investigation of T1 -weighted OE-MRI to identify regional tumor hypoxia. The quantification of AUCOE has translational potential as a clinical biomarker of hypoxia.

Exploring intra- and inter-reader variability in uni-dimensional, bi-dimensional, and volumetric measurements of solid tumors on CT scans reconstructed at different slice intervals.

OBJECTIVE: Understanding magnitudes of variability when measuring tumor size may be valuable in improving detection of tumor change and thus evaluating tumor response to therapy in clinical trials and care. Our study explored intra- and inter-reader variability of tumor uni-dimensional (1D), bi-dimensional (2D), and volumetric (VOL) measurements using manual and computer-aided methods (CAM) on CT scans reconstructed at different slice intervals. MATERIALS AND METHODS: Raw CT data from 30 patients enrolled in oncology clinical trials was reconstructed at 5, 2.5, and 1.25 mm slice intervals. 118 lesions in the lungs, liver, and lymph nodes were analyzed. For each lesion, two independent radiologists manually and, separately, using computer software, measured the maximum diameter (1D), maximum perpendicular diameter, and volume (CAM only). One of them blindly repeated the measurements. Intra- and inter-reader variability for the manual method and CAM were analyzed using linear mixed-effects models and Bland-Altman method. RESULTS: For the three slice intervals, the maximum coefficients of variation for manual intra-/inter-reader variability were 6.9%/9.0% (1D) and 12.3%/18.0% (2D), and for CAM were 5.4%/9.3% (1D), 11.3%/18.8% (2D) and 9.3%/18.0% (VOL). Maximal 95% reference ranges for the percentage difference in intra-reader measurements for manual 1D and 2D, and CAM VOL were (-15.5%, 25.8%), (-27.1%, 51.6%), and (-22.3%, 33.6%), respectively. CONCLUSIONS: Variability in measuring the diameter and volume of solid tumors, manually and by CAM, is affected by CT slice interval. The 2.5mm slice interval provides the least measurement variability. Among the three techniques, 2D has the greatest measurement variability compared to 1D and 3D.

Age at onset and vascular pathology in late-life depression.

OBJECTIVE: There is considerable evidence to suggest that late-onset depression may be etiologically distinct from early-onset depression. The aim of this study was to compare vascular function and magnetic resonance imaging-defined brain ischemic changes between early-onset depressed (EOD) and late-onset depressed (LOD) subjects. DESIGN: Case-control study. PARTICIPANTS: Twenty-five subjects with late-life depression recruited from secondary care were divided into groups with EOD (<60 years, 11 subjects) and LOD (>60 years, 14 subjects). MEASURES: All subjects underwent a variety of vascular assessments including pulse wave analysis, pulse wave velocity, carotid intima media thickness (IMT), and magnetic resonance imaging of the brain to assess white matter hyperintensities. RESULTS: The mean age of LOD subjects was 71.3 ± 4.0 years and EOD was 73.6 ± 4.7 years (p = NS). There were no baseline differences in vascular risk or sociodemographic variables. LOD subjects had significantly higher common carotid IMT (EOD: 0.06 [0.01]; LOD: 0.09 [0.02], p = 0.02), carotid plaques (EOD: 2.1 [1.1]; LOD: 5.4 [3.9], p = 0.02), and peripheral augmentation index (EOD: 81.7 [7.9]; LOD: 96.2 [21.6], p = 0.04) when compared with early-onset subjects, indicating more vascular pathology. There were no group differences in white matter hyperintensities. Age at onset of depression was positively correlated with peripheral augmentation index, common carotid IMT, and plaque index. CONCLUSION: This study suggests that elderly subjects with LOD have greater vascular impairment than those with an early-onset illness. Whether preventing vascular disease at an earlier age may decrease the risk of last onset depression is a potential area for future research.

Vascular function in older adults with depressive disorder.

BACKGROUND: Cerebrovascular disease plays an important role in depressive disorder, especially in older adults. An understanding of vascular function in depression is important etiologically and to develop innovative treatments that may improve prognosis by ameliorating vascular damage. METHODS: This study assessed endothelial function, arterial stiffness, and atherosclerosis in a variety of vessel beds in 25 elderly subjects with depressive disorder compared with 21 nondepressed control subjects. Subjects underwent pulse wave velocity, pulse wave analysis, carotid intima media thickness analysis, and magnetic resonance imaging. A subset (16 patients and 15 control subjects) had assessment of biopsied small artery dilatation to acetylcholine to further assess endothelial function. RESULTS: The mean sample age was 72.4 years with an average age at onset for depression of 60 years. Mean carotid intima media thickness was significantly higher in depressed subjects (p < .01). Pulse wave velocity was 1.6 m/sec higher in depressed subjects (borderline significance). There was a significant reduction in the dilatation response to acetylcholine in preconstricted small arteries (p = .01). On magnetic resonance imaging, depressed subjects had significantly more dilated Virchow-Robin spaces in the basal ganglia (p = .01). Depressed subjects had greater volume of white matter lesions in all regions, but this did not reach statistical significance. There were no baseline differences in vascular risk. CONCLUSIONS: Depression in the elderly is associated with poorer endothelial function and more atherosclerosis. This is associated with a greater white matter hyperintensities lesion load and basal ganglia microangiopathy. The use of vasoprotective drugs to improve endothelial function or retard atherosclerosis as depression-modifying agents should be explored.

Robust tissue boundary detection for cerebral cortical thickness estimation.

This paper presents an algorithm for determining regional cerebral grey matter cortical thickness from magnetic resonance scans. In particular, the modification of a gradient-based edge detector into an iso-grey-level boundary detector for reliably determining the low-contrast grey-white matter interface is described and discussed. The reproducibility of the algorithm over 31 gyral regions is assessed using repeat scans of four subjects, and a technique for correcting the misplacement of the grey-white matter boundary is shown to significantly reduce the systematic error on the reproducibility.