The definitive diagnosis of acute lymphoblastic leukemia (ALL), as a highly prevalent cancer, requires invasive, expensive, and time-consuming diagnostic tests. ALL diagnosis using peripheral blood smear (PBS) images plays a vital role in the initial screening of cancer from non-cancer cases. The examination of these PBS images by laboratory users is riddled with problems such as diagnostic error because the nonspecific nature of ALL signs and symptoms often leads to misdiagnosis. Herein, a model based on deep convolutional neural networks (CNNs) is proposed to detect ALL from hematogone cases and then determine ALL subtypes. In this paper, we build a publicly available ALL data set, comprised 3562 PBS images from 89 patients suspected of ALL, including 25 healthy individuals with a benign diagnosis (hematogone) and 64 patients with a definitive diagnosis of ALL subtypes. After color thresholding-based segmentation in the HSV color space by designing a two-channel network, 10 well-known CNN architectures (EfficientNet, MobileNetV3, VGG-19, Xception, InceptionV3, ResNet50V2, VGG-16, NASNetLarge, InceptionResNetV2, and DenseNet201) were employed for feature extraction of different data classes. Of these 10 models, DenseNet201 achieved the best performance in diagnosis and classification. Finally, a model was developed and proposed based on this state-of-the-art technology. This deep learning-based model attained an accuracy, sensitivity, and specificity of 99.85, 99.52, and 99.89%, respectively. The proposed method may help to distinguish ALL from benign cases. This model is also able to assist hematologists and laboratory personnel in diagnosing ALL subtypes and thus determining the treatment protocol associated with these subtypes. The proposed data set is available at https://www.kaggle.com/mehradaria/leukemia and the implementation (source code) of proposed method is made publicly available at https://github.com/MehradAria/ALL-Subtype-Classification.

中文翻译：

---

基于外周血涂片图像的 B-ALL 诊断及其亚型分类的快速高效 CNN 模型

急性淋巴细胞白血病 (ALL) 作为一种高度流行的癌症，其明确诊断需要侵入性、昂贵且耗时的诊断测试。使用外周血涂片 (PBS) 图像进行 ALL 诊断在从非癌症病例中初步筛查癌症中起着至关重要的作用。实验室用户对这些 PBS 图像的检查充满了诸如诊断错误之类的问题，因为 ALL 体征和症状的非特异性通常会导致误诊。在此，提出了一种基于深度卷积神经网络 (CNN) 的模型来检测血细胞中的 ALL，然后确定 ALL 亚型。在本文中，我们构建了一个公开可用的 ALL 数据集，包括来自 89 名疑似 ALL 患者的 3562 张 PBS 图像，包括 25 名诊断为良性（血红素）的健康个体和 64 名明确诊断为 ALL 亚型的患者。通过设计双通道网络在 HSV 颜色空间中进行基于颜色阈值的分割后，10 种著名的 CNN 架构（EfficientNet、MobileNetV3、VGG-19、Xception、InceptionV3、ResNet50V2、VGG-16、NASNetLarge、InceptionResNetV2 和 DenseNet201 ) 被用于不同数据类的特征提取。在这 10 个模型中，DenseNet201 在诊断和分类方面取得了最好的表现。最后，基于这种最先进的技术开发并提出了一个模型。这种基于深度学习的模型分别达到了 99.85%、99.52% 和 99.89% 的准确性、敏感性和特异性。所提出的方法可能有助于区分 ALL 和良性病例。该模型还能够帮助血液学家和实验室人员诊断 ALL 亚型，从而确定与这些亚型相关的治疗方案。提议的数据集可在 https://www.kaggle.com/mehradaria/leukemia 获得，提议方法的实现（源代码）在 https://github.com/MehradAria/ALL-Subtype-Classification 公开.