Impact of rescanning and normalization on convolutional neural network performance in multi-center, whole-slide classification of prostate cancer

Abstract

Algorithms can improve the objectivity and efficiency of histopathologic slide analysis. In this paper, we investigated the impact of scanning systems (scanners) and cycle-GAn-based normalization on algorithm performance, by comparing different deep learning models to automatically detect prostate cancer in whole-slide images. Specifically, we compare U-Net, DenseNet and EfficientNet. Models were developed on a multi-center cohort with 582 WSIs and subsequently evaluated on two independent test sets including 85 and 50 WSIs, respectively, to show the robustness of the proposed method to differing staining protocols and scanner types. We also investigated the application of normalization as a pre-processing step by two techniques, the whole-slide image color standardizer (WSICS) algorithm, and a cycle-GAN based method. For the two independent datasets we obtained an AUC of 0.92 and 0.83 respectively. After rescanning the AUC improves to 0.91/0.88 and after style normalization to 0.98/0.97. In the future our algorithm could be used to automatically pre-screen prostate biopsies to alleviate the workload of pathologists. Prostate cancer is the most common cancer in men and the third most common tumor type worldwide 1,2. In 2018, 1.3 million new cases have been diagnosed (7.1% of all diagnosed cancers), and 28% of these patients died as a result of the disease 1. Prostate cancer is typically diagnosed through ultrasound-guided biopsy after initial suspicion has arisen through, for example, a prostate specific antigen (PSA) blood test. During the prostate biopsy procedure, 6-12 core samples are taken from a patient 3 resulting worldwide in more than 15 million specimens annually, which is expected to increase further with the aging of the population. All these specimens have to be evaluated by pathologists. However, in many countries there is a lack of pathologists which is only expected to increase in the years to come. Automating (part of) the evaluation of prostate biopsies might help mitigate the lack of clinical pathology. The histopathological analysis could be streamlined significantly if these negative slides (i.e. slides without pathology) could automatically be excluded without expelling any slides containing cancer. Significant progress has been made in this respect, revealing the huge potential of deep learning (DL) methods 4-6. In histopathology, deep learning based algorithms have been used to solve a variety of tasks, such as mitotic Figure detection 7 , lung adenocarcinoma segmentation 4 , glomeruli detection 8 or tissue analysis in colorectal cancer 9. However, histological slides from different institutions show heterogeneous appearance as a result of the different preparation and staining procedures (different colors, intensity, saturation) (Fig. 1). As a result, there is a high probability that a model trained on data from one medical center may not be applicable to slides from another center. The key challenge is to develop a system robust to a variety of biological, staining or scanning settings.

Quality control stress test for deep learning-based diagnostic model in digital pathology

“…1 ). Additional layer of heterogeneity are color schemes of WSI scanners from different manufactures which imprint every scanned slide [ 35 , 36 ]. The same slides from one institution digitized by three different scanners (Datasets 3–5) in our study are highly heterogenous in terms of color scheme, brightness and contrast (Supplementary Fig.…”

Deep Learning Approaches and Applications in Toxicologic Histopathology: Current Status and Future Perspectives

“…Similar work has been done on brain tissue that contains lesional (e.g., lymphomas) and nonlesional(e.g., normal cortical gray and white matter) categories where the unsupervised model generates multiple clusters representing different areas. [ 54 ] Furthermore, these lower-dimensional features in unsupervised learning can be used to artificially generate new images (e.g., stain normalization[ 55 ] and stain transfer). Figure 5 shows how these two different tasks use unsupervised architectures.…”

Normalization of HE-stained histological images using cycle consistent generative adversarial networks

“…Recent publications investigated in the use of deep learning approaches with Generative Adversarial Networks (GANs) and showed the benefits compared to the conventional methods [ 9 , 10 ]. It was also shown how normalizing images using GANs can highly improve results of image classification [ 11 ] or segmentation [ 12 ]. Mahapatra et al [ 13 ] integrate self-supervised semantic information such as geometric and structural patterns at different layers to improve stain normalization with CycleGANs.…”