水泡状葡萄胎（HM）是常见的妊娠滋养细胞疾病之一，具有恶性潜力。组织病理学检查是诊断HM的主要方法。然而，由于HM的病理学特征模糊而混乱，在病理学家之间存在显著的观察者差异，导致临床实践中的误诊和过诊。有效特征提取可以显著提高诊断过程的准确性和速度。深度神经网络（DNN）已经被证明具有优秀的特征提取和分割能力，在许多其他疾病的临床实践中得到广泛应用。我们构建了一种基于深度学习的CAD方法，在显微镜下实时识别HM水泡病变。为解决由于从HM镜片图像中提取有效特征困难而导致的病变分割挑战，我们提出了一种使用DeepLabv3+的水泡病变识别模块，采用我们的新复合损失函数和逐步训练策略，以在像素和病变水平上识别水泡病变表现出更好的性能。同时，还开发了基于傅里叶变换的图像拼接模块和图像序列边缘扩展模块，使识别模型更适用于临床实践中移动幻灯片的情况。这种方法还解决了模型在图像边缘识别方面表现不佳的情况。我们使用广泛采用的DNN在HM数据集上评估了我们的方法，并选择DeepLabv3+作为分割模型。对比实验表明，边缘扩展模块能够在像素级IoU和病变级IoU方面将模型性能提高最多3.4%和9.0%。至于最终结果，我们的方法能够实现像素级IoU为77.0％，精度为86.0％，病变级召回率为86.2％，响应时间为每帧82ms。实验表明，我们的方法能够在实时跟踪幻灯片的移动下，将精确标记的HM水泡病变全面展示于显微视野中。据我们所知，这是第一种在HM病变识别中利用深度神经网络的方法。该方法通过强大的特征提取和分割能力，提供了一种稳健准确的解决方案，对HM的辅助诊断有帮助。版权所有©2023 Elsevier B.V.发表。

Hydatidiform mole (HM) is one of the most common gestational trophoblastic diseases with malignant potential. Histopathological examination is the primary method for diagnosing HM. However, due to the obscure and confusing pathology features of HM, significant observer variability exists among pathologists, leading to over- and misdiagnosis in clinical practice. Efficient feature extraction can significantly improve the accuracy and speed of the diagnostic process. Deep neural network (DNN) has been proven to have excellent feature extraction and segmentation capabilities, which is widely used in clinical practice for many other diseases. We constructed a deep learning-based CAD method to recognize HM hydrops lesions under the microscopic view in real-time.To solve the challenge of lesion segmentation due to difficulties in extracting effective features from HM slide images, we proposed a hydrops lesion recognition module that employs DeepLabv3+ with our novel compound loss function and a stepwise training strategy to achieve great performance in recognizing hydrops lesions at both pixel and lesion level. Meanwhile, a Fourier transform-based image mosaic module and an edge extension module for image sequences were developed to make the recognition model more applicable to the case of moving slides in clinical practice. Such an approach also addresses the situation where the model has poor results for image edge recognition.We evaluated our method using widely adopted DNNs on an HM dataset and chose DeepLabv3+ with our compound loss function as the segmentation model. The comparison experiments show that the edge extension module is able to improve the model performance by at most 3.4% regarding pixel-level IoU and 9.0% regarding lesion-level IoU. As for the final result, our method is able to achieve a pixel-level IoU of 77.0%, a precision of 86.0%, and a lesion-level recall of 86.2% while having a response time of 82 ms per frame. Experiments show that our method is able to display the full microscopic view with accurately labeled HM hydrops lesions following the movement of slides in real-time.To the best of our knowledge, this is the first method to utilize deep neural networks in HM lesion recognition. This method provides a robust and accurate solution with powerful feature extraction and segmentation capabilities for auxiliary diagnosis of HM.Copyright © 2023. Published by Elsevier B.V.