Laboratorio de Pesquisa Medica - LIM55, Divisao de Urologia, Faculdade de Medicina FMUSP, Universidade de Sao Paulo, Sao Paulo, SP, BR.
Clinics (Sao Paulo). 2021 Oct 29;76:e3198. doi: 10.6061/clinics/2021/e3198. eCollection 2021.
This study aims to evaluate the ability of deep learning algorithms to detect and grade prostate cancer (PCa) in radical prostatectomy specimens.
We selected 12 whole-slide images of radical prostatectomy specimens. These images were divided into patches, and then, analyzed and annotated. The annotated areas were categorized as follows: stroma, normal glands, and Gleason patterns 3, 4, and 5. Two analyses were performed: i) a categorical image classification method that labels each image as benign or as Gleason 3, Gleason 4, or Gleason 5, and ii) a scanning method in which distinct areas representative of benign and different Gleason patterns are delineated and labeled separately by a pathologist. The Inception v3 Convolutional Neural Network architecture was used in categorical model training, and a Mask Region-based Convolutional Neural Network was used to train the scanning method. After training, we selected three new whole-slide images that were not used during the training to evaluate the model as our test dataset. The analysis results of the images using deep learning algorithms were compared with those obtained by the pathologists.
In the categorical classification method, the trained model obtained a validation accuracy of 94.1% during training; however, the concordance with our expert uropathologists in the test dataset was only 44%. With the image-scanning method, our model demonstrated a validation accuracy of 91.2%. When the test images were used, the concordance between the deep learning method and uropathologists was 89%.
Deep learning algorithms have a high potential for use in the diagnosis and grading of PCa. Scanning methods are likely to be superior to simple classification methods.
本研究旨在评估深度学习算法在前列腺根治性切除标本中检测和分级前列腺癌（PCa）的能力。
我们选择了 12 张前列腺根治性切除标本的全切片图像。这些图像被划分为小块，然后进行分析和注释。注释区域分为基质、正常腺体和 Gleason 模式 3、4 和 5。进行了两种分析：i）一种分类图像分类方法，将每张图像标记为良性或 Gleason 3、Gleason 4 或 Gleason 5，ii）一种扫描方法，由病理学家分别划定和标记代表良性和不同 Gleason 模式的不同区域。在分类模型训练中使用了 Inception v3 卷积神经网络架构，在扫描方法中使用了 Mask Region-based 卷积神经网络进行训练。在训练后，我们选择了三张未在训练中使用的新全切片图像作为我们的测试数据集来评估模型。使用深度学习算法对图像进行分析的结果与病理学家的结果进行了比较。
在分类方法中，训练后的模型在训练过程中获得了 94.1%的验证准确率；然而，在测试数据集与我们的专家泌尿病理学家的一致性仅为 44%。使用图像扫描方法，我们的模型验证准确率为 91.2%。当使用测试图像时，深度学习方法与泌尿病理学家之间的一致性为 89%。
深度学习算法在前列腺癌的诊断和分级中有很高的应用潜力。扫描方法可能优于简单的分类方法。
