IHC has become the workhorse of molecular phenotyping for tissues and currently serves as a reliable surrogate to actually performing expensive molecular testing. However, IHC is time-consuming, can be expensive, and dependent on appropriate tissue handling procedures, reagents, and expert laboratory technicians. Furthermore, immunostain findings require visual inspection using a microscope and thus depend on the subjective interpretation of pathologists [21, 22]. Recent technological progress in digital pathology and AI has shown that these new modalities can be used to not only improve efficiency of pathologists, but also provide comparable diagnostic accuracy to pathologists employing traditional light microscopy [12]. Within the domain of surgical pathology, AI-based algorithms can analyze digitized histomorphologic features to effectively distinguish neoplastic and non-neoplastic lesions [23, 24], detect metastasis in lymph nodes [25], predict genomic fusion status within renal neoplasms [26], subtype renal tumors [1], detect prostate cancer in biopsy material [27], as well as grade aggressiveness of certain tumors [4]. To date, AI-based studies have been applied to prostate cancer pathology to assist with diagnosis, Gleason grading, prognosis, as as well as predict underlying molecular aberrations such as phosphatase and tensin homology (PTEN) loss [2, 28, 29]. In the present study, we developed a deep learning model to predict ERG rearrangement status in patients with prostatic adenocarcinoma. To the best of our knowledge, this is the first study to identify this genomic status directly from scanned H&E-stained slides in surgically resected prostate cancer cases.

TMPRSS2ERG rearrangement, the most common ETS gene fusion in prostate cancer, brings ERG expression under androgen control via androgen receptor-mediated TMPRSS2 regulation and results in over-expression of ERG protein [3]. Microscopically, while some ERG rearranged prostate cancers are enriched with features such as intraluminal blue mucin and prominent nucleoli, the spectrum of morphology is quite variable and inconsistently predictive of the presence of an ERG rearrangement at the genomic level [7]. Hence, it remains challenging to faithfully distinguish prostate cancer with ERG rearrangement from those with wild type ERG and other molecular subtypes based only on the microscopic evaluation of H&E-stained pathological tissues. As a result, ERG gene rearrangement status is usually confirmed by immunohistochemical identification of the overexpression of ERG protein or by dual-color break-apart FISH. However, these ancillary tests require additional time and resources, and they consume precious tissue.

In this study, we demonstrated that a digitized H&E-stained slide analyzed using a deep learning-based model can successfully predict ERG fusion status in the majority of prostate cancer cases. We believe that this algorithm can eliminate the need for using extra tumor tissue to perform lengthy and expensive ancillary studies to assess for ERG gene rearrangement in prostatic adenocarcinoma. Our AI model was able to accurately predict the presence of an ERG gene rearrangement in a large number of cases with varying morphologic patterns and Grade groups, including tumors with low-grade (Grade Group 2 or less) and high-grade features (Grade Group 3 or higher). Higher accuracy was seen in lower-grade tumors. One possibility for this observation may be that higher-grade tumors typically exhibit more diverse morphology.

ERG gene rearrangement is known to contribute to the pathogenesis of prostate cancer and provides important clues about the multifocality and metastatic dissemination of this disease. The specificity of this gene rearrangement in prostate cancer allows ERG evaluation by IHC to be of diagnostic value in both primary and metastatic tumors originating from the prostate [6, 30, 31]. TMPRSS2ERG fusions are also prime candidates for the development of new diagnostic assays, including urine-based noninvasive assays [32]. There have been conflicting reports regarding the prognostic value of ERG gene rearrangement and its overexpression in prostatic cancer. Hägglöf et al. have demonstrated that high expression of ERG is associated with higher Gleason score, aggressive disease and poor survival rates [33]. Similarly, Nam et al. demonstrated that the TMPRSS2-ERG fusion gene predicts cancer recurrence after surgical treatment and that this prediction is independent of grade, stage and prostate specific antigen (PSA) levels in blood [34]. Mehra et al. demonstrated ERG rearrangement to be associated with a higher stage in prostate cancer [35]. A subsequent study by Fine et al. demonstrated a subset of prostate cancers with TMPRSS2-ERG copy number increase, with or without rearrangement, to be associated with higher Gleason score [36]. Nevertheless, other studies have found no association between TMPRSS2-ERG fusion and stage, grade, recurrence, or progression [37, 38]. Additionally, in the TCGA dataset, in our limited analyses, we did not see any other particular genetic mutation that is significantly different between ERG rearranged and non-ERG rearranged cases. Currently, there is no ERG-targeted therapy approved for treatment of prostate cancer. However, peptidomimetic targeting of transcription factor fusion products has been demonstrated to provide a promising therapeutic strategy for prostate cancer [39]. Previous ERG fusion driven biomarker clinical trials utilized interrogation of ERG rearrangement status employing IHC or FISH tests [40]. Our study provides a viable and inexpensive alternative to ancillary tissue-based testing methods to detect ERG rearrangement status in prostate cancer.

Our study has several strengths and potential limitations. Notable strengths include the use of H&E stained slides only (without the need for concurrent genomic investigation) to predict ERG gene fusion status in prostate cancer, utilization of WSI, and employment of diverse datasets including in-house and TCGA datasets with different H&E staining qualities to improve the robustness of our algorithm. This application carries strength in eliminating the need for complex molecular testing utilizing FISH, next-generation sequencing, or molecular surrogate assays like immunohistochemistry; utilizing of H&E slides only allows an easy, economical and efficient methodology to detect ERG gene rearrangement utilizing AI developed model. Importantly, our study paves a foundation for utilizing basic laboratory tools in assessing genomic rearrangements in diverse set of human malignancies (of prostate and other genitourinary tumors).

Computational limitations for both training and test set evaluation were considered when deciding which neural network architecture to utilize. Commonly used architectures for image classification tasks include Inception, VGG16, ResNet50, and MobileNet, among others. Each architecture comes with its own strengths and limitations and each one is designed to be optimal under specific circumstances. For example, the Inception architecture serves the purpose of reducing computational cost by implementing a shallower network compared to ResNet50 which may negatively impact computational accuracy. MobileNetV2, part of the MobileNet family, further addresses issues of size and speed and is optimally designed for mobile device applications which often require computationally limited platforms. This is accomplished by utilizing 19 inverted residual bottleneck layers following the initial fully convolutional layer, The 19 bottleneck layers are subsequently followed by a point convolutional layer, pooling average layer, and a final convolutional layer. Taking into consideration that these AI-applications are ultimately intended for clinical laboratory settings which may not have access to high-end computational hardware, we ultimately chose a MobileNetV2 architecture pre-trained on ImageNet as our base model due to its balance between accuracy and computational cost [41, 42].

Hardware limitations necessitating relatively small input tiles may contribute to our model’s performance. Our training set was relatively enriched in lower-grade tumors as high-grade cancers are less common in daily clinical urological practice. Follow-up studies incorporating more higher-grade tumors will be needed to better assess the performance of our AI-based tool in such scenario. Our algorithm was developed using resection specimens, and further studies would be needed to interrogate findings in biopsy specimens that display smaller volumes of tumor; as a consequence, cut-offs used in this study may need to be adjusted. For the purposes of this study, we did not address disease heterogeneity and multifocality; future studies are likely to address these phenomena.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.


This article is autogenerated using RSS feeds and has not been created or edited by OA JF.

Click here for Source link (https://www.biomedcentral.com/)

By admin