Bungum, Aaron O.

AbstractBackgroundObtaining a diagnosis and treating pulmonary malignancies during the same anesthesia requires either an on‐site pathologist or a system for remotely evaluating microscopic images. Cytology specimens are challenging to remotely assess given the need to navigate through dispersed and three‐dimensional cell clusters. Remote navigation is possible using robotic telepathology, but data are limited on the ease of use of current systems, particularly for pulmonary cytology.MethodsAir dried modified Wright‐Giemsa stained slides from 26 touch preparations of transbronchial biopsies and 27 smears of endobronchial ultrasound guided fine needle aspirations were scored for ease of adequacy assessment and ease of diagnosis on robotic (rmtConnect Microscope) and non‐robotic telecytology platforms. Diagnostic classifications were compared between glass slides and the robotic and non‐robotic telecytology assessments.ResultsCompared to non‐robotic telecytology, robotic telecytology had a greater ease of adequacy assessment and non‐inferior ease of diagnosis. The median time to diagnosis using robotic telecytology was 85 s (range 28–190 s). Diagnostic categories were concordant for 76% of cases in robotic versus non‐robotic telecytology and 78% of cases in robotic telecytology versus glass slide diagnosis. Weighted Cohen's kappa scores for agreement in these comparisons were 0.84 and 0.72, respectively.ConclusionsUse of a remote‐controlled robotic microscope improved the ease of adequacy assessment compared to non‐robotic telecytology and enabled strongly concordant diagnoses to be expediently rendered. This study provides evidence that modern robotic telecytology is a feasible and user‐friendly method of remotely and potentially intraoperatively rendering adequacy assessments and diagnoses on bronchoscopic cytology specimens.

2616 Background: Non-small cell lung cancer (NSCLC) tissue is a valuable resource for diagnosis, treatment planning, and drug development. Current 2D histopathological techniques introduce under-sampling error (i.e., a single 5 µm section represents 0.5% of a 1 mm thick biopsy), interobserver variability, and failure to capture the biology contained within the entire tissue sample. We have developed a suite of technologies to stain, chemically clarify, image, visualize, and analyze entire, intact lung tissue samples. Herein, we present analysis methods for comprehensive 3D assessment of intact NSCLC tissue. Methods: Human NSCLC tissue, stored frozen in optimal cutting temperature (OCT) compound, was fixed in 4% paraformaldehyde, stained with nuclear (TOPRO-3) and general protein (Eosin) fluorescent dyes, and optically cleared using a modified iDISCO protocol with ethyl cinnamate as the refractive index matching solution. Twenty tissue samples measuring 1-5 mm3 in volume were imaged at 2 µm/pixel resolution with a hybrid open-top light-sheet microscope. Smaller regions of interest (ROIs) with interesting pathologic features were re-imaged at higher resolution, 0.17 µm/pixel. Visualization and analysis was performed using Aivia software, custom spatial analysis software and python scripts. Results: All NSCLC tissue samples were successfully imaged in 3D at both low and high resolution. Computational approaches, such as machine learning, were used to segment and classify tumor cells and lymphocytes within the high resolution ROIs. In one example we quantified a 0.0045 mm3 ROI containing 1x109 voxels. We segmented 1,474 unique objects, and classified 683 (46%) of the objects as lymphocytes and 309 (20%) of the cells as tumor cells. We found that on average the lymphocytes were within 12 µm of the nearest tumor cell. Additionally the tissue was broken into separate lymphocyte dense (58% of volume) and tumor dense regions (42% of volume). Further analysis is in progress to quantify more features, such as tertiary lymphoid structures, stromal cells, and fibrosis over the full 20 sample dataset. Conclusions: We successfully implemented 3D machine learning analysis pipelines using images from intact NSCLC tissue samples measuring up to 5 mm3. The imaging was performed using our custom hybrid open-top light-sheet microscope. This novel method enables us to identify and quantify key histological features such as tumor cells and lymphocytes within the tumor microenvironment. In addition, machine learning driven algorithms clustered cells into distinct tissue regions in 3D which enables a more precise assessment of disease state and heterogeneity. Further investigation is ongoing to link these results to patient outcomes.

8549 Background: First-line treatment of early stage non-small cell lung cancer (NSCLC) is surgical resection, but with a 5-year survival of only 54%, rates of future relapse are high. Identifying patients at high risk of relapse can help guide adjuvant treatment decisions. Deep convolutional neural network (DCNN) AI models trained on tumor histology have shown incredible flexibility as potential biomarkers, both in lung cancer and more generally across many malignancies. While DCNN models can obtain extremely accurate results when used for routine purposes such as diagnosis, subtyping, and grading of malignancies, most models trained for prognostication or treatment response prediction do not reach performance sufficient for clinical application. Uncertainty quantification (UQ) – a family of techniques that give DCNN models the ability to report confidence alongside predictions – is an underexplored avenue in cancer AI that may help further improve performance and clinical application of models designed to provide clinicians with estimations of risk. Methods: To explore the potential use of UQ in clinically-oriented DCNN models, we trained models on a single-institution, retrospective digital tumor histology slide cohort from patients with Stage I-III NSCLC who underwent surgical resection to predict risk of future relapse. Estimation of uncertainty was performed using dropout as a Bayesian approximation, and uncertainty thresholds were calculated from training sets using a novel method to identify and remove low-confidence predictions. For comparison, a separate multivariate logistic regression was trained in cross-validation using known clinical risk factors. Results: We trained DCNN models on slides from 198 patients (40 relapsed, 158 without relapse). In this cohort, 130 patients had stage I disease (65.6%), 42 had stage II (21.2%), and 26 had stage III (13.2%). The average age was 69 years, 85% were current or previous smokers, and 38 received guideline-concordant adjuvant chemotherapy. Without UQ, a DCNN model predicted risk of future relapse with an average area under receiver operator curve (AUROC) of 0.74 across three-fold cross-validation. Using UQ estimation, 63% of slides were reported with high confidence. Relapse prediction was significantly improved in the high-confidence cohort, with an average AUROC of 0.83 in cross-validation. With a specificity of 70%, this corresponds to an average sensitivity of 86.8% across the three cross-folds (79.1%, 85.5%, and 95.7%). In comparison, a clinical-only multivariate regression model predicted relapse with a cross-validated AUROC of 0.67. Conclusions: This method of uncertainty quantification appears to be a powerful tool to predict lung cancer recurrence risk from digital histopathology while simultaneously providing clinicians with a measurement of algorithm trustworthiness.