Portal de Periódicos da CAPES

The Diagnosis of Chronic Myeloid Leukemia with Deep Adversarial Learning

2022; Elsevier BV; Volume: 192; Issue: 7 Linguagem: Inglês

10.1016/j.ajpath.2022.03.016

1525-2191

Zelin Zhang, Xianqi Huang, Qi Yan, Yani Lin, Enbin Liu, Yingchang Mi, Shi Liang, Hao Wang, Jun Xu, Kun Ru,

Cytokine Signaling Pathways and Interactions

Chronic myeloid leukemia (CML) is a clonal proliferative disorder of granulocytic lineage, with morphologic evaluation as the first step for a definite diagnosis. This study developed a conditional generative adversarial network (cGAN)–based model, CMLcGAN, to segment megakaryocytes from myeloid cells in bone marrow biopsies. After segmentation, the statistical characteristics of two types of cells were extracted and compared between patients and controls. At the segmentation phase, the CMLcGAN was evaluated on 517 images (512 × 512) which achieved a mean pixel accuracy of 95.1%, a mean intersection over union of 71.2%, and a mean Dice coefficient of 81.8%. In addition, the CMLcGAN was compared with seven other available deep learning–based segmentation models and achieved a better segmentation performance. At the clinical validation phase, a series of seven-dimensional statistical features from various cells were extracted. Using the t-test, five-dimensional features were selected as the clinical prediction feature set. Finally, the model iterated 100 times using threefold cross-validation on whole slide images (58 CML cases and 31 healthy cases), and the final best AUC was 84.93%. In conclusion, a CMLcGAN model was established for multiclass segmentation of bone marrow cells that performed better than other deep learning–based segmentation models. Chronic myeloid leukemia (CML) is a clonal proliferative disorder of granulocytic lineage, with morphologic evaluation as the first step for a definite diagnosis. This study developed a conditional generative adversarial network (cGAN)–based model, CMLcGAN, to segment megakaryocytes from myeloid cells in bone marrow biopsies. After segmentation, the statistical characteristics of two types of cells were extracted and compared between patients and controls. At the segmentation phase, the CMLcGAN was evaluated on 517 images (512 × 512) which achieved a mean pixel accuracy of 95.1%, a mean intersection over union of 71.2%, and a mean Dice coefficient of 81.8%. In addition, the CMLcGAN was compared with seven other available deep learning–based segmentation models and achieved a better segmentation performance. At the clinical validation phase, a series of seven-dimensional statistical features from various cells were extracted. Using the t-test, five-dimensional features were selected as the clinical prediction feature set. Finally, the model iterated 100 times using threefold cross-validation on whole slide images (58 CML cases and 31 healthy cases), and the final best AUC was 84.93%. In conclusion, a CMLcGAN model was established for multiclass segmentation of bone marrow cells that performed better than other deep learning–based segmentation models. Myeloproliferative neoplasm (MPN) is a group of chronic hematopoietic stem cell tumors that arise from myeloid lineage.1Gotlib J. World Health Organization-defined eosinophilic disorders: 2017 update on diagnosis, risk stratification, and management.Am J Hematol. 2017; 92: 1243-1259Crossref PubMed Scopus (136) Google Scholar As a prototype, chronic myeloid leukemia (CML) was initially described in 1845 which accounted for approximately 15% of leukemia cases worldwide. Males have a mildly predominant incidence of CML at any age.2Bennett J.H. Case of hypertrophy of the spleen and liver, in which death took place from suppuration of the blood.Edinb Med Surg J. 1845; 64: 413-423Google Scholar, 3Gunnarsson R. Rosenquist R. New insights into the pathobiology of chronic lymphocytic leukemia.J Hematopathol. 2011; 4: 149-163Crossref Scopus (2) Google Scholar, 4Swerdlow S.H. Campo E. Harris N.L. Jaffe E.S. Pileri S.A. Stein H. Thiele J. Vardiman J.W. WHO classification of tumours of haematopoietic and lymphoid tissues.Rev 4th ed. International Agency for Research on Cancer, Lyon, France2008Google Scholar The characteristic genetic feature of CML is the Philadelphia chromosomal translocation between chromosomes 9 and 22, resulting in the BCR-ABL1 fusion gene.5Rowley J.D. A new consistent chromosomal abnormality in chronic myelogenous leukaemia identified by quinacrine fluorescence and giemsa staining.Nature. 1973; 243: 290-293Crossref PubMed Scopus (3429) Google Scholar This translocation causes constitutive activation of the tyrosine kinase ABL1 and promotes cell growth and survival via augmenting signals from downstream pathways.6Chereda B. Melo J.V. Natural course and biology of CML.Ann Hematol. 2015; 94: 107-121Crossref PubMed Scopus (150) Google Scholar CML can exhibit as chronic phase, accelerated phase, or blast phase7Vardiman J.W. Chronic myelogenous leukemia, BCR-ABL1+.Am J Clin Pathol. 2009; 2: 250Crossref Scopus (48) Google Scholar based on clinical and laboratory manifestations, such as an elevated white blood cell count, increased eosinophils or basophils, as well as >20% of blast cells.8Goldman J. Marin D. Olavarria E. Apperley J.F. Clinical decisions for chronic myeloid leukemia in the imatinib era.Semin Hematology. 2003; 40: 98-103Crossref PubMed Scopus (31) Google Scholar Tyrosine kinase inhibitors are the frontline treatment of CML. Specifically tailored tyrosine kinase inhibitor therapy and precise molecular monitoring can significantly repress the progression from the chronic phase to the accelerated phase or blast phase and increase the 10-year survival rate for patients with CML.9Croes S. Vries F.D. Drug–drug interactions with tyrosine-kinase inhibitors.Lancet Oncol. 2014; 15: e315-e326Abstract Full Text Full Text PDF PubMed Scopus (200) Google Scholar The diagnosis of CML relies on integrated analysis of the results of various laboratory tests, such as morphologic, genetic, and molecular biological testing. Although the BCR-ABL1 fusion gene is the disease's hallmark, histologic and cytologic tests are still the first steps to approach the disorder, from both bone marrow and peripheral blood with typical morphologic features.8Goldman J. Marin D. Olavarria E. Apperley J.F. Clinical decisions for chronic myeloid leukemia in the imatinib era.Semin Hematology. 2003; 40: 98-103Crossref PubMed Scopus (31) Google Scholar However, the morphologic examination is time-consuming and subject to individual bias. A reliable automated counter has yet to be explored given the intrinsic complexity of bone marrow components. The increase in machine learning algorithms combined with digital pathology sheds light on this dilemma. Some image processing–based methods have been applied to blood cell segmentation tasks.10Bouchet A. Montes S. Ballarin V. Díaz I. Intuitionistic fuzzy set and fuzzy mathematical morphology applied to color leukocytes segmentation.Signal Image Video Process. 2020; 14: 1-9Crossref Scopus (21) Google Scholar, 11Dorini Minetto L. Semiautomatic white blood cell segmentation based on multiscale analysis.IEEE J Biomed Health Inform. 2013; 6: 1-7Google Scholar, 12Li Y. Zhu R. Mi L. Cao Y. Yao D. Segmentation of white blood cell from acute lymphoblastic leukemia images using dual-threshold method.Comput Math Methods Med. 2016; 2016: 9514707Crossref PubMed Scopus (91) Google Scholar Deep convolutional neural networks were used in various medical imaging processing tasks, including cell or nuclei segmentation,13Gu Z. Cheng J. Fu H. Zhou K. Hao H. Zhao Y. Zhang T. Gao S. Liu J. CE-Net: context encoder network for 2D medical image segmentation.IEEE Trans Med Imaging. 2019; : 1-12Google Scholar, 14Lu Y. Fan H. Li Z. Leukocyte segmentation via end-to-end learning of deep convolutional neural networks.Int Conf Intell Sci Big Data Eng. 2019; 11935: 191-200Google Scholar, 15Zhang M. Li X. Xu M. Li Q. Automated semantic segmentation of red blood cells for sickle cell disease.IEEE J Biomed Health Inform. 2020; 3: 76Google Scholar detection,16Jung C. Abuhamad M. Alikhanov J. Mohaisen A. Han K. Nyang D.H. W-Net: a CNN-based architecture for white blood cells image classification.arXiv. 2019; ([Preprint] doi:)10.48550/arXiv.1910.01091Google Scholar, 17Qiu W. Guo J. Li X. Xu M. Zhang M. Guo N. Li Q. Multi-label detection and classification of red blood cells in microscopic images.arXiv. 2019; ([Preprint] doi:)10.48550/arXiv.1910.02672Google Scholar, 18Xia T. Jiang R. Fu Y.Q. Jin N. Automated blood cell detection and counting via deep learning for microfluidic point-of-care medical devices.IOP Conf Ser Mater Sci Eng. 2019; 646: 012048Crossref Scopus (17) Google Scholar classification,16Jung C. Abuhamad M. Alikhanov J. Mohaisen A. Han K. Nyang D.H. W-Net: a CNN-based architecture for white blood cells image classification.arXiv. 2019; ([Preprint] doi:)10.48550/arXiv.1910.01091Google Scholar and tissue-type segmentation.19Xu J. Luo X. Wang G. Gilmore H. Madabhushi A. A deep convolutional neural network for segmenting and classifying epithelial and stromal regions in histopathological images.Neurocomputing. 2016; 191: 214-223Crossref PubMed Scopus (352) Google Scholar This study presents an end-to-end model, CMLcGAN, based on conditional generative adversarial networks (cGANs)20Mirza M. Osindero S. Conditional generative adversarial nets.arXiv. 2014; ([Preprint] doi:)10.48550/arXiv.1411.1784Google Scholar for multiclass bone marrow cell segmentation. The proposed architecture integrates an end-to-end segmentation network and deep adversarial learning. In the CMLcGAN, the primary role of adversarial learning is to provide a trainable loss function to guide the segmentation network for a better training at the deep feature level. The study consisted of 89 patients, including 58 patients with CML and 31 controls from the SINO-US Diagnostics Lab. The protocols were reviewed and approved by the institutional review board of SINO-US Diagnostics Lab. Formalin-fixed, paraffin-embedded bone marrow biopsy specimens were stained with hematoxylin and eosin (H&E). The slides were scanned at 40× objective using a Pannoramic 250 Flash III Dx scanner (3DHISTECH, Budapest, Hungary) to generate the whole slide image (WSI). Eighty-nine H&E–stained WSIs were used for the proposed whole pipeline of fully automated auxiliary diagnostic verification. For each WSI, the background area was removed by preprocessing, and the effective area was automatically tiled into 2048 × 2048 images for clinical verification. The framework of the automatic diagnosis of CML is shown in Figure 1. The H&E–stained slides were photographed at 400× optical resolution. The main steps of the whole pipeline were as below. First, preprocess for each WSI was performed to remove the background region and block the effective regions into 2048 × 2048 size images as the clinical diagnosis data set.21Riasatian A. Rasoolijaberi M. Babaei M. Tizhoosh H.R. A comparative study of U-Net topologies for background removal in histopathology images.arXiv. 2020; ([Preprint] doi:)10.48550/arXiv.2006.06531Google Scholar Second, the pretrained CMLcGAN was used to segment megakaryocytes (MKs). The blue mask represented the MKs, and the white mask represented the myeloid cells. Third, seven statistical features were extracted. A statistical standardization method was used to avoid statistical differences caused by different actual areas among slices. Specifically, all statistical features for each WSI were divided by the number of image blocks as the standardized values. The first five-dimensional features were selected as the diagnostic prediction features. Fourth, the five selected features were used to make the final diagnosis. A series of binary classifiers on a data set from 89 cases were used to perform a standard 200 threefold cross-validation, and the final cross-validation means ± SEM AUC value of the best classifier was 84.93% ± 1.1%. In the medical image segmentation, the UNet22Ronneberger O. Fischer P. Brox T. U-net: convolutional networks for biomedical image segmentation.arXiv. 2015; ([Preprint] doi:)10.48550/arXiv.1505.04597Google Scholar architecture for deep learning provided an accurate image semantic segmentation location. The deformable convolution layer realized the free-form deformation of the feature learning process, resulting in high segmentation and classification accuracy. Compared with the general encoder-decoder segmentation models, UNet added skip connections between the encoder feature maps and the decoder feature maps at the same scale. The skip connection was an effective way to maintain most image details. Zhou et al23Zhou Z. Siddiquee M.M.R. Tajbakhsh N. Liang J. UNet++: A Nested U-Net Architecture for Medical Image Segmentation.arXiv. 2018; ([Preprint] doi:)10.48550/arXiv.1807.10165Google Scholar found that the segmentation performance of the UNet++ network was better than that of the UNet network. UNet++ was inspired by DenseNet24Huang G. Liu Z. Van Der Maaten L. Weinberger K.Q. Densely connected convolutional networks.2017 IEEE Con Comput Vis Pattern Recognit. 2017; : 2261-2269Crossref Scopus (20220) Google Scholar and increased the amount of gradient calculation of the network by adding dense connections between modules to improve the network's segmentation performance. The specific model structure is shown in Figure 2. Compared with UNet, the main improvement of UNet++ was the multiscale decoding branches. All the encoder and decoder deep features for the same scale were cascaded on the channel dimension using skip connections. A deep-supervised learning method was used to allow multiple decoders to output the segmentation results so that the network could be pruned. Given the advantages of the UNet++ dense gradient connections, UNet++ was used as the generator network of CMLcGAN. GANs25Goodfellow I. Pouget-Abadie J. Mirza M. Xu B. Warde-Farley D. Ozair S. Courville A. Bengio Y. Generative adversarial networks.Adv Neural Inf Process Syst. 2014; 3: 2672-2680Google Scholar were proposed in 2014 to tackle the problem of image generation, and consisted of a generator network and a discriminator network. The generator network generated fake images that approximated the real images, whereas the discriminator network classified the generated images from the real images to construct an adversarial loop. GAN effectively used the discriminator to discriminate the output of the generative model at the deep feature level, equivalent to using a trainable loss function that provided a better guidance to the training of the generator. Some medical image–related studies have achieved better performance by incorporating GANs.26Quiros A.C. Murray-Smith R. Yuan K. PathologyGAN: learning deep representations of cancer tissue.arXiv. 2019; ([Preprint] doi:)10.48550/arXiv.1907.02644Google Scholar, 27Gupta L. Klinkhammer B.M. Boor P. Merhof D. Gadermayr M. GAN-based image enrichment in digital pathology boosts segmentation accuracy//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2019: 631-639Google Scholar, 28Yang T. Wu T. Li L. Zhu C. SUD-GAN: deep convolution generative adversarial network combined with short connection and dense block for retinal vessel segmentation.J Digital Imaging. 2020; 33: 946-957Crossref PubMed Scopus (41) Google Scholar The cGAN added a control condition vector to the generator and discriminator of GAN to control the generator's output. The objective function of cGAN was as follows:minGmaxDL(D,G)=Ex∼Pdata(x)[logD(x|y)]+Ez∼Pz(z)[log(1−D(G(z|y)))](1) where G and D represent the generator and the discriminator, respectively, x represents the real images of random input, z represents the random input noises, and y represents the conditional control vectors. Pix2pix29Isola P. Zhu J.Y. Zhou T. Efros A. Image-to-image translation with conditional adversarial networks.IEEE Conf Comput Vis Pattern Recognit. 2017; : 5967-5976Google Scholar was an image translation model based on cGAN for translating images in different fields. The conditional adversarial objective of Pix2pix was as follows:minGmaxDL(D,G)=Ex,y[logD(x,y)]+Exz[log(1−D(x,G(x,z)))](2) where x and y represent the input images and target images, respectively, and z represents the random input noises. Compared with cGAN, the condition input of Pix2pix was the input images x, and the generator did not require conditional input. To enhance the generating ability of the generator, Pix2pix also added a regularized L130De Bot K. Gommans P. Rossing C. L1 loss in an L2 environment: Dutch immigrants in France. First Language attrition, 1991: 87-98Google Scholar loss function, and the L1 loss function was as follows:L1(G)=Ex,y,z[y−G(x,z)∥1](3) where x, y, and z represent generator input images, target images, and random input noises, respectively. A CML automatic diagnostic system was built, including a multiclass module for bone marrow cell segmentation and a diagnostic verification module. A series of comparative experiments showed that the proposed CMLcGAN had a satisfactory segmentation performance of multiclass bone marrow cells by achieving a mean pixel accuracy of 95.1%, a mean intersection over union (IoU) of 71.2%, and a mean Dice coefficient of 81.8%. After the segmentation, five statistical features were selected, and a standard 100-repeats threefold cross-validation31Browne M.W. Cross-validation methods.J Math Psycholog. 2000; 44: 108-132Crossref PubMed Scopus (726) Google Scholar was performed. By verifying the CML diagnostic system from the data set of 89 cases, the final means ± SEM AUC was determined to be 84.93% ± 0.011. The detailed structure of the proposed CMLcGAN is shown in Figure 2. The generator used the UNet++ L23Zhou Z. Siddiquee M.M.R. Tajbakhsh N. Liang J. UNet++: A Nested U-Net Architecture for Medical Image Segmentation.arXiv. 2018; ([Preprint] doi:)10.48550/arXiv.1807.10165Google Scholar structure. All convolution modules adopted the structure of dense convolutional blocks. All upsampling used 2× bilinear interpolation, and all downsampling used 2× maximum pooling layer. The skip connections were used to merge the encoder's deep feature maps with the decoder features of the same scale in the channel dimension. The discriminator of CMLcGAN was a binary classifier that contained four 2× convolutional downsampling layers. A single convolutional layer was used as the feature output layer of the discriminator, and the mean value of the output feature was used as the probability value of the final adversarial loss. The parameters of the four convolutional layers were 3 × 3 convolution kernels, 2 × 2 convolution stride, and 1 × 1 zero paddings. Beyond the four convolutional layers, the batch normalization layer32Ioffe S. Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift.arXiV. 2015; ([Preprint] doi:)10.48550/arXiv.1502.03167Google Scholar was used for feature standardization, and a LeakyReLU33Dubey A. Jain V. Comparative study of convolution neural network's relu and leaky-relu activation functions. Springer, New York, NY2019: 873-880Google Scholar activation function was used before the next convolutional layer. The single convolution output layer used a 3 × 3 convolution kernel, which only changed the number of feature output channels to 1. The feature output visualization of the discriminator is shown in Figure 3. The optimization object of CMLcGAN was similar to Pix2pix without increasing the diversity of the generator's results in the multiclass task for bone marrow cell segmentation. Therefore, the generative adversarial loss of CMLcGAN was as follows:minGmaxDL(D,G)=Ex,y[logD(x,y)]+Ex[log(1−D(x,G(x)))](4) where the L1 loss function was used in the loss function of CMLcGAN to enhance the generation ability of the UNet++ generator, as shown in Formula 3. Therefore, the ultimate object of CMLcGAN was as follows:minGmaxDL(D,G)+λL1(5) where λ = 10. It controlled the weight of L1 loss. Performance evaluation included local and global evaluation of the semantic segmentation results of multiclass bone marrow cells. Confusion matrix was calculated between the segmentation results of segmentation models on the independent test set and the target ground truth, as well as the pixel-level true-positive (TP), true-negative (TN), false-positive (FP), and false-negative (FN) results. Pixel accuracy (PA), intersection ratio (IoU), and Dice coefficient were used to demonstrate the advantages of CMLcGAN. PA was calculated as the ratio of the number of correctly defined pixels in the model's output to the number of all pixels in the image. PA was defined as follows: PA=TP+TNTP+TN+FP+FN(6) IoU measured the overlap rate between model segmentation results and target ground truth. IoU was defined as follows:IoU=TPTP+FP+FN(7) The Dice coefficient measured the similarity between the model output and target ground truth and was used in the evaluation of semantic segmentation models. The Dice coefficient was defined as follows:DICE=2×TP2×TP+FP+FN(8) The evaluation results of CMLcGAN's multiclass of bone marrow cells are shown in Figure 4. As a benchmark, the mean ± SEM PA of CMLcGAN segmentation results from bone marrow cells was 95.1% ± 0.015%, the mean ± SD IoU was 71.2% ± 0.091%, and the mean ± SEM Dice coefficient was 81.8% ± 0.081%. The means of PA, IoU, and Dice coefficient of multiclass of bone marrow cell segmentation were used as global evaluation indicators to compare and verify CMLcGAN and the other seven semantic segmentation models. The overall comparison and evaluation results are given in Table 1. It appears that the CMLcGAN had the best performance.Table 1Comparison of Evaluation Results of CMLcGAN and Seven Other ModelsMethodPA, mean ± SEMIoU, mean ± SEMDice coefficient, mean ± SEMMKsMyeloid cellsMKsMyeloid cellsMKsMyeloid cellsFCN8s0.823 ± 0.0360.958 ± 0.0040.267 ± 0.0420.851 ± 0.01410.399 ± 0.0520.915 ± 0.009FCN16s0.819 ± 0.0480.955 ± 0.0030.266 ± 0.0540.841 ± 0.0120.396 ± 0.0670.909 ± 0.007FCN32s0.758 ± 0.0680.952 ± 0.0040.203 ± 0.0510.831 ± 0.0160.318 ± 0.0650.903 ± 0.011SegNet0.869 ± 0.0350.957 ± 0.0030.338 ± 0.0580.845 ± 0.0140.479 ± 0.0650.911 ± 0.008Pix2pix0.819 ± 0.1230.927 ± 0.0050.301 ± 0.1190.745 ± 0.0210.432 ± 0.1440.851 ± 0.013Unet0.953 ± 0.0110.961 ± 0.0050.577 ± 0.0460.861 ± 0.0210.711 ± 0.0310.921 ± 0.012Unet++0.927 ± 0.0310.959 ± 0.0020.492 ± 0.1160.851 ± 0.0090.634 ± 0.1040.916 ± 0.006CMLcGAN0.959 ± 0.0170.962 ± 0.0070.631 ± 0.0750.893 ± 0.0230.754 ± 0.0660.922 ± 0.014CMLcGAN, chronic myeloid leukemia conditional generative adversarial network; IoU, intersection over union; MKs, megakaryocytes; PA, pixel accuracy. Open table in a new tab CMLcGAN, chronic myeloid leukemia conditional generative adversarial network; IoU, intersection over union; MKs, megakaryocytes; PA, pixel accuracy. In the training phase, Adam was used as the optimizer.34Sweke R. Wilde F. Meyer J. Schuld M. Eiser J. Stochastic gradient descent for hybrid quantum-classical optimization.Quantum. 2020; 4: 314Crossref Scopus (107) Google Scholar The initial learning rate was 0.0002, β1 was 0.9, and β2 was 0.99. The model was implemented in Python software version 3.7 (Python Software Foundation, Fredericksburg, VA; https://www.python.org) and Pytorch software version 1.2.0 (Facebook's AI Research lab, Menlo Park, CA; https://pytorch.org). A total of 16 minibatches and 100 training epochs were used during training with the Nvidia GPU 1080ti with cudnn version 7.6 and Intel CPU Core (TM) i7-4790@3.60GHz. CMLcGAN directly performed end-to-end segmentation without any preprocessing or postprocessing tasks to train samples. The test results on a WSI with the pretrained CMLcGAN are shown in Figure 5. The proposed CMLcGAN was evaluated on 517 H&E–stained patches with a pixel size of 512 × 512. The patches were extracted from 89 WSIs. All images were analyzed by two licensed pathologists (K.R. and E.L.) independently, and two types of bone marrow cells were annotated (MKs labeled in blue and myeloid cells labeled in white). The model was trained and validated with 10 iterations using threefold cross-validation. In the testing phase, the standard fivefold cross-validation was used to evaluate and compare the performance of CMLcGAN. The original data set was randomly divided into five sub–data sets with the same number of samples. Four of the sub–data sets were used as the training set, and the remaining subset was used as an independent test set to perform the model training and testing. Five models were trained with the same training epoch, and the mean of the model's test evaluation results from five independent test sets was used as the final evaluation index. Through cross-validation, all samples in the data set were used in training and testing to eliminate the unstable model performance caused by sample diversity. The segmentation results (Figure 5) appeared acceptable to the pathologists. To verify the segmentation performance of CMLcGAN, CMLcGAN was also compared with seven other end-to-end segmentation models based on deep learning: FCN8s, FCN16s, FCN32s,35Han C. Duan Y. Tao X. Lu J. Dense convolutional networks for semantic segmentation.IEEE Access. 2019; 7: 43369-43382Crossref Scopus (28) Google Scholar UNet,22Ronneberger O. Fischer P. Brox T. U-net: convolutional networks for biomedical image segmentation.arXiv. 2015; ([Preprint] doi:)10.48550/arXiv.1505.04597Google Scholar SegNet,36Badrinarayanan V. Kendall A. Cipolla R. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation.IEEE Trans Pattern Anal Mach Intell. 2017; 39: 2481-2495Crossref PubMed Scopus (10261) Google Scholar UNet++,23Zhou Z. Siddiquee M.M.R. Tajbakhsh N. Liang J. UNet++: A Nested U-Net Architecture for Medical Image Segmentation.arXiv. 2018; ([Preprint] doi:)10.48550/arXiv.1807.10165Google Scholar and Pix2pix.29Isola P. Zhu J.Y. Zhou T. Efros A. Image-to-image translation with conditional adversarial networks.IEEE Conf Comput Vis Pattern Recognit. 2017; : 5967-5976Google Scholar On the basis of the multitype segmentation results, the seven-dimensional statistical features of the multiclass of bone marrow cells were constructed: the number of MKs, the number of myeloid cells, the density of MKs, and the four parameters of the MK size (maximum, minimum, mean, and SD).37Bartels P. Olson G. Computer Analysis of Lymphocyte Images. Springer US, 1980Crossref Google Scholar,38Mendelsohn M. Kolman W.A. Perry B. Prewitt J.M. Morphological analysis of cells and chromosomes by digital computer.Methods Inf Med. 1965; 04: 163-167Crossref Scopus (6) Google Scholar The size of the MKs was measured with the radius of equal area circles. Standard t-test was used to perform feature testing on the extracted seven-dimensional statistical features. The P values of the seven-dimensional features are given in Table 2. Seven features were extracted, where the standard t-test was used to perform feature testing. The first five dimensions of the smallest P value were selected as the clinical prediction indexes.Table 2P Values of the Extracted FeaturesStatistical featurePMKs, n4.796 × 10−6Myeloid cells, n1.996 × 10−3MK density3.332 × 10−4MK size, mean ± SD6.128 × 10−1 ± 6.156 × 10−2Maximum MK size3.178 × 10−3Minimum MK size6.199 × 10−4MK, megakaryocyte. Open table in a new tab MK, megakaryocyte. On the basis of the P value39Hosmer D. Lemeshow S. Confidence interval estimation of interaction.Epidemiology (Cambridge, Mass). 1992; 3: 452-456Crossref PubMed Scopus (911) Google Scholar of the t-test, five dimensional features were chosen for further analysis: the number of MKs, the number of myeloid cells, the density of the MKs, and the maximum and minimum size of the MKs. The boxplots of the selected five-dimensional features are shown in Figure 6. The selected five-dimensional features were used to make the clinical prediction for 89 cases. The standard 100 iterations with threefold cross-validation were adopted to obtain stable validation results. At the same time, linear discriminant analysis classifier,40Balakrishnama S. Ganapathiraju A. Linear discriminant analysis-a brief tutorial. Institute for Signal and information Processing, Mississippi State, MS1998: 1-8Google Scholar random forest classifier,41Ho T.K. Random decision forests//Proceedings of 3rd international conference on document analysis and recognition.IEEE. 1995; 1: 278-282Google Scholar neural network classifier,42Rogova G.L. Combining the results of several neural network classifiers. Elsevier Science Ltd, Philadelphia, PA2008Crossref Scopus (15) Google Scholar linear support vector machine classifier,43Chen P.H. Lin C.J. Schölkopf B. A tutorial on ν-support vector machines.Appl Stochastic Models Business Industry. 2005; 21: 111-136Crossref Scopus (394) Google Scholar decision tree classifier44Du W. Zhan Z. Building decision tree classifier on private data.Proc IEEE Int Conf Privacy Security Data Mining Australian Computer Society Inc. 2002; 14: 1-8Google Scholar AdaBoost classifier,45Kearns M. Valiant L. Cryptographic limitations on learning Boolean formulae and finite automata.J Am Coll Med. 1994; 41: 67-95Google Scholar quadratic discriminant analysis classifier,46Srivastava S. Gupta M.R. Frigyik B.A. Bayesian quadratic discriminant analysis.J Machine Learn Res. 2007; 8: 1277-1305Google Scholar and nearest neighbors classifier47Horton P. Nakai K. Better prediction of protein cellular localization sites with the k nearest neighbors classifier.Proc Int Conf Intelligent Syst Molec Biol. 1997; 5: 147-152PubMed Google Scholar were used in parallel for cross-validation. The receiver operating characteristic48McClish D. Analyzing a portion of the ROC curve, medical decision making. 9. 1989: 190-195Google Scholar curve of the cross-validation result is shown in Figure 7. The linear support vector machine classifier with the superlative AUC (84.93% ± 0.011%) had the best results. The best-performing binary classifier was the linear support vector machine classifier with a mean ± SEM achieved AUC49Lobo J.M. Jiménez-Valverde A. Real R. AUC: a misleading measure of the performance of predictive distribution models.Global Ecol Biogeography. 2008; 17: 145-151Crossref Scopus (2296) Google Scholar of 84.93% ± 0.011%. The proposed CMLcGAN and diagnostic prediction feature set appeared to deliver satisfactory performance in CML diagnosis. The morphologic evaluation is still the cornerstone of the diagnosis and classification of hematologic disorders. With the development of deep learning algorithms and digital pathology, artificial intelligence appears to be a promising way for health organizations and medical practitioners to diagnose and classify hematologic disorders. The MKs have varied but characteristic morphologic findings in MPNs and usually appear smaller than their normal counterparts in CML. The study found that the proposed CMLcGAN had a high diagnostic accuracy to recognize the atypical MKs in the bone marrow biopsies of CML, indicating a promising way to assist hematopathologists. The five-dimensional features include the number of MKs, the number of myeloid cells, the density of MKs, and the maximum and minimum size of MKs. The boxplots of those features reveal that the tissue of CML displays a conspicuous increase in the number of MKs, the number of myeloid cells, and the density of MKs, equivalent to the hypercellularity and increased MKs in the bone marrow. In addition, the maximum and minimum of MKs are more sizable and smaller, perhaps because of observation bias, sampling issues, algorithm imperfections, or segmentation errors. A CMLcGAN model for multiclass segmentation of bone marrow cells was developed, and a CML automatic diagnostic system was constructed. This model achieves a better segmentation performance than other deep learning–based segmentation models. On the basis of the segmentation results, the newly constructed diagnostic prediction feature set effectively differentiates patients with CML from healthy controls. Herein, novel artificial intelligence system is presented using sufficient data sets and effective morphologic parameters to verify the diagnostic prediction feature set. Our next goal is to establish the MPN artificial intelligence diagnostic standard data set by expanding the samples from CML to other MPNs. The establishment of the data sets from various MPNs would make the CMLcGAN model closer to the practical application.