Faculty Bibliography

Background: Artificial intelligence (AI) has been increasingly used in surgical pathology to assist pathologists for diagnosis, grading and staging purposes. We describe in this report a proof of concept study that pathologists themselves could use a commercial deep learning software to develop an AI model to assist the classification of flat urothelial lesions.
Design(s): 227 flat urothelial lesions including normal (N) (66), benign reactive (BR) (53), dysplasia (D) (49) and carcinoma in situ (CIS) (59) from one institution (TMC) were reviewed, and representative H&E micrographs of taken at 10X magnification by pathologists on various cameras attached to microscopes. Fifty % of these images were imported into Cognex VisionPro Deep Learning 2.0 software tool, which was trained to develop a flat urothelial classifier (Model 1) to distinguish between N, BR, D and CIS. The classifier comprised of a chain of 3 deep learning models to first remove the blank space in the images, to identify urothelial regions and finally to classify the urothelial regions. An F-score was used to measure the accuracy of the model to classify the images of flat urothelial lesions. 191 micrographs of flat urothelial lesions from a different institution (NYU), including N (35), BR (52), D (47) and CIS (57), were used as validation cases. Fifty % of latter images were used to train a new model (Model 2) based on the Model 1.
Result(s): The F-score for Model 1 on TMC cases was 96.1%. 97.1%, 85.2%, 100% and 100% of cases in N, BR, D and CIS were correctly classified. When Model 1 was used on NYU cases, the overall F-score was 44.9%. Model 1 was retrained using 50% of NYU images to generate Model 2. The F-score for Model 2 on NYU images improved to 86.8%, while it was 88.9% on TMC images.
Conclusion(s): This proof of concept study confirms that a pathologist with little background in deep learning software development can use a commercially available deep learning software to develop a robust AI model to assist the classification of flat urothelial lesions. While a model developed using one set of images from one lab may not be directly applied to images from a different lab, it could nevertheless be retrained using the images from the second lab to achieve high classification accuracy. Studies are underway to develop AI models for flat urothelial classification on whole slide images

Importance/UNASSIGNED:There is clinical equipoise for COVID-19 convalescent plasma (CCP) use in patients hospitalized with COVID-19. Objective/UNASSIGNED:To determine the safety and efficacy of CCP compared with placebo in hospitalized patients with COVID-19 receiving noninvasive supplemental oxygen. Design, Setting, and Participants/UNASSIGNED:CONTAIN COVID-19, a randomized, double-blind, placebo-controlled trial of CCP in hospitalized adults with COVID-19, was conducted at 21 US hospitals from April 17, 2020, to March 15, 2021. The trial enrolled 941 participants who were hospitalized for 3 or less days or presented 7 or less days after symptom onset and required noninvasive oxygen supplementation. Interventions/UNASSIGNED:A unit of approximately 250 mL of CCP or equivalent volume of placebo (normal saline). Main Outcomes and Measures/UNASSIGNED:The primary outcome was participant scores on the 11-point World Health Organization (WHO) Ordinal Scale for Clinical Improvement on day 14 after randomization; the secondary outcome was WHO scores determined on day 28. Subgroups were analyzed with respect to age, baseline WHO score, concomitant medications, symptom duration, CCP SARS-CoV-2 titer, baseline SARS-CoV-2 serostatus, and enrollment quarter. Outcomes were analyzed using a bayesian proportional cumulative odds model. Efficacy of CCP was defined as a cumulative adjusted odds ratio (cOR) less than 1 and a clinically meaningful effect as cOR less than 0.8. Results/UNASSIGNED:Of 941 participants randomized (473 to placebo and 468 to CCP), 556 were men (59.1%); median age was 63 years (IQR, 52-73); 373 (39.6%) were Hispanic and 132 (14.0%) were non-Hispanic Black. The cOR for the primary outcome adjusted for site, baseline risk, WHO score, age, sex, and symptom duration was 0.94 (95% credible interval [CrI], 0.75-1.18) with posterior probability (P[cOR<1] = 72%); the cOR for the secondary adjusted outcome was 0.92 (95% CrI, 0.74-1.16; P[cOR<1] = 76%). Exploratory subgroup analyses suggested heterogeneity of treatment effect: at day 28, cORs were 0.72 (95% CrI, 0.46-1.13; P[cOR<1] = 93%) for participants enrolled in April-June 2020 and 0.65 (95% CrI, 0.41 to 1.02; P[cOR<1] = 97%) for those not receiving remdesivir and not receiving corticosteroids at randomization. Median CCP SARS-CoV-2 neutralizing titer used in April to June 2020 was 1:175 (IQR, 76-379). Any adverse events (excluding transfusion reactions) were reported for 39 (8.2%) placebo recipients and 44 (9.4%) CCP recipients (P = .57). Transfusion reactions occurred in 2 (0.4) placebo recipients and 8 (1.7) CCP recipients (P = .06). Conclusions and Relevance/UNASSIGNED:In this trial, CCP did not meet the prespecified primary and secondary outcomes for CCP efficacy. However, high-titer CCP may have benefited participants early in the pandemic when remdesivir and corticosteroids were not in use. Trial Registration/UNASSIGNED:ClinicalTrials.gov Identifier: NCT04364737.

Introduction & Objectives: Delay between prostate biopsy (PB) and pathologic diagnosis leads to a concern of inadequate sampling and repeated biopsy. Stimulated Raman Histology (SRH) is a novel microscopic technique allowing real time, label-free, high-resolution microscopic images of unprocessed, un-sectioned tissue. We evaluated the accuracy of pathologist interpretation of PB SRH as compared to traditional hematoxylin and eosin (H&E) stained slides.
Material(s) and Method(s): Men undergoing prostatectomy were included in an IRB approved prospective study. 18-gauge PB cores, taken ex vivo from prostatectomy specimen, were scanned in a SRH microscope at 20 microns depth over 10-14 minutes using two Raman shifts: 2845cm-1 and 2930cm-1, to create SRH images. The cores were then processed as per normal pathologic protocols. 16 PB containing benign/prostate cancer histology were used as a SRH training cohort for 4 GU pathologists (1, 3, 2x >15 yrs experience), who were then tested on a set of 32 PB imaged by SRH and processed by traditional H&E. Sensitivity, specificity, and concordance for PCa detection on SRH relative to a consensus H&E were assessed. With a two-sided alpha level of 5%, it was calculated 32 SRH imaged biopsies would provide 90% power to detect concordance (k).
Result(s): PB cores were imaged in 2-3 separate strips (11-21 minutes) shown in Figure 1. In identifying any cancer, pathologists achieved moderate concordance (k=0.570; p<0.001) which improved when identifying GGG 2-5 PCa only (k=0.640, p<0001; sensitivity 96.4%; specificity 58.3%). In predicting Gleason score, the concordance for each pathologist varied from poor to moderate (k range -0.163 to 0.457). After individual assessment was completed a pathology consensus conference was held for the interpretation of the SRH PB. In identifying any prostate cancer, pathologists achieved near perfect concordance (k=0.925; p<0.001; sensitivity 95.6%; specificity 100%). When evaluating SRH in a consensus conference, the group prediction of Gleason score improved to moderate concordance (k=0.470; p<0.001). (Figure Presented) (Figure Presented)
Conclusion(s): SRH produces high quality microscopic images that allow for accurate identification of PCa in real-time without need for sectioning or tissue-processing. Individual pathologist performance varied highly suggesting potential for improvement with further training. Future SRH interpretation by convolutional neural network may further enhance GGG prediction
Copyright

When multiple cores are biopsied from a single magnetic resonance imaging (MRI)-targeted lesion, Gleason grade may be assigned for each core separately or for all cores of the lesion in aggregate. Because of the potential for disparate grades, an optimal method for pathology reporting MRI lesion grade awaits validation. We examined our institutional experience on the concordance of biopsy grade with subsequent radical prostatectomy (RP) grade of targeted lesions when grade is determined on individual versus aggregate core basis. For 317 patients (with 367 lesions) who underwent MRI-targeted biopsy followed by RP, targeted lesion grade was assigned as (1) global Grade Group (GG), aggregated positive cores; (2) highest GG (highest grade in single biopsy core); and (3) largest volume GG (grade in the core with longest cancer linear length). The 3 biopsy grades were compared (equivalence, upgrade, or downgrade) with the final grade of the lesion in the RP, using κ and weighted κ coefficients. The biopsy global, highest, and largest GGs were the same as the final RP GG in 73%, 68%, 62% cases, respectively (weighted κ: 0.77, 0.79, and 0.71). For cases where the targeted lesion biopsy grade scores differed from each other when assigned by global, highest, and largest GG, the concordance with the targeted lesion RP GG was 69%, 52%, 31% for biopsy global, highest, and largest GGs tumors (weighted κ: 0.65, 0.68, 0.59). Overall, global, highest, and largest GG of the targeted biopsy show substantial agreement with RP-targeted lesion GG, however targeted global GG yields slightly better agreement than either targeted highest or largest GG. This becomes more apparent in nearly one third of cases when each of the 3 targeted lesion level biopsy scores differ. These results support the use of global (aggregate) GG for reporting of MRI lesion-targeted biopsies, while further validations are awaited.

Prostate cancer is the most frequently diagnosed cancer in males and its development and progression remains an important area of study. Recently, long non-coding RNAs (lncRNAs) have been evidenced as key players in cancer pathogenesis. Specifically, dysregulation of long non-coding RNA (lncRNA) expression has shown to affect tumor proliferation and metastasis, acting as either tumor suppressors or oncogenes. However, its specific mechanisms and functions in prostate cancer remain unclear. This review provides an overview of currently available information on prostate cancer-related lncRNAs, including GAS5, GAS-007, MEG3, PCA3, PCAT14, PCAT1, PVT1, UCA1, SChLAP1, MALAT1, HOTAIR, and NEAT1. Notable tumor growth inhibitors include GAS5 and MEG3. GAS5 is evidenced to interfere with the AKT/MTOR signaling pathway through targeting microRNA mir-103. MEG3, however, is proposed to inhibit the cycle, sponge miR-9-5p, and induce gene silencing. PCAT1, PVT1, and UCA1 are important tumor growth promoters. PCAT1 is indicated to be a transcriptional repressor, a mir-145-5P sponge, and a P13K/AKT pathway activator. Studies suggest that PVT1 acts via microRNA targeting and regulating proliferating cell nuclear antigen. UCA1 may sponge miR-204 and miR-331-3p as well as regulate myosin VI. Thorough understanding of these lncRNAs may elucidate new aspects of prostate cancer pathology and serve a pivotal role in developing novel diagnostic and prognostic techniques.

Prostate cancer (PCa) incidence and mortality rate vary among racial and ethnic groups with the highest occurrence in African American (AA) men who have mortality rates twice that of Caucasians (CA). In this study, we focused on differential expression of proteins in AA prostate cancer compared to CA using Protein Pathway Array Analysis (PPAA), in order to identify protein biomarkers associated with PCa racial disparity. Fresh frozen prostate samples (n=90) obtained from radical prostatectomy specimens with PCa, including 25 AA tumor, 21 AA benign, 23 CA tumor, 21 CA benign samples were analyzed. A total of 286 proteins and phosphoproteins were assessed using PPAA. By PPAA analysis, 33 proteins were found to be significantly differentially expressed in tumor tissue (n=48, including both CA and AA) in comparison to benign tissue (n=42). We further compared protein expression levels between AA and CA tumor groups and found that 3 proteins were differentially expressed (P<0.05 and q<5%). Aurora was found to be significantly increased in AA tumors, while Cyclin D1 and HNF-3a proteins were downregulated in AA tumors. Predicted risk score was significantly different between AA and CA ethnic groups using logistic regression analysis. In conclusion, we identified Aurora, Cyclin D1 and HNF-3a proteins as being differentially expressed between AA and CA in PCa tissue. Our study suggests that these proteins might be involved in different pathways that lead to aggressive PCa behavior in AA patients, potentially serving as biomarkers for the PCa racial disparity.