Open Access
Add to BibSonomy
Abstract
Multiple myeloma (MM) is a type of blood cancer where plasma cells abnormally multiply and crowd out regular blood cells in the bones. Automated analysis of bone marrow smear examination is considered promising to improve the performance and reduce the labor cost in MM diagnosis. To address the drawbacks in established methods, which mainly aim at identifying monoclonal plasma cells (monoclonal PCs) via binary classification, in this work, considering that monoclonal PCs is not the only basis in MM diagnosis, for the first we construct a multi-object detection model for MM diagnosis. The experimental results show that our model can handle the images at a throughput of 80 slides/s and identify six lineages of bone marrow cells with an average accuracy of 90.8%. This work makes a step further toward full-automatic and high-efficiency MM diagnosis.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Multiple myeloma (MM) is a hematological malignancy characterized by uncontrolled clonal proliferation of plasma cells (PCs) in bone marrow (BM) [1–3] and represents the second most common hematological malignancy across the globe [4]. Currently, MM is more common in the elderly and is considered as incurable disease. According to the guidelines of the international myeloma working group [5], the diagnosis of MM is determined by the basic CRAB symptoms (including hypercalcemia, renal insufficiency, anemia, and bone lesions) as well as various laboratory examinations. The most important of which are the proportion of monoclonal PCs, the detection of serum-free light chains, and the detection of lesions by MRI [6–8]. Although recent studies have proposed different methods for the detection of MM, including the microscopic view of BM biopsy tissue morphology, immunophenotyping techniques (flow cytometry and immunohistochemistry), fluorescence in situ hybridization of PCs, and electrophoresis for the detection of monoclonal immunoglobulin [9]. Morphological detection of PCs still plays an indispensable role in the diagnosis and efficacy evaluation of MM. Nevertheless, BM cytology examination involves complex cell types, relying on manual microscopy, there are a variety of problems, such as long time-consuming, low efficiency, high labor costs, and poor reproducibility, which greatly affect the screening of hematological tumor and treatment.
In addition, manual cytology also meets serious problems, such as a lack of professional and technical employees, as well as a heavy workload, which adversely affects the MM diagnosis, and the accurate diagnosis of the disease strongly depends on the medical experts’ abilities and experiences. Although computer-aided diagnosis technology has made certain developments and been applied in medical analysis in recent years, such as gastric cancer detection [10], peripheral blood analysis [11], etc. Furthermore, BM of myeloma patients often contains a large number of monoclonal immunoglobulins and serum-free light chains, which severely affects the coloring of the specimens after Wright-Giemsa staining, leading to poor computer recognition results [12]. Additionally, the proportion of monoclonal PCs fluctuates greatly in MM, and the erythroid, granulocytic, lymphoid, monocytic cells and others in the BM will disrupt monoclonal PCs identification [13,14]. Existing methods mainly focus on two classifications, that is, they mainly detect monoclonal PCs and others. Nevertheless, myeloma patients can present with a variety of clinical symptoms, such as infection symptoms, anemia, etc., all of which affect the granulocyte and erythroid cell proportions. Therefore, the detection of common cell types plays an auxiliary role in the accurate diagnosis of MM. Besides this, traditional methods are usually multi-stage analyses that cannot be used efficiently and rapidly in the clinic. Moreover, the clinic also cares about the proliferation of cells of other lineages in the BM, which is conducive to the discovery of suppressed cell lineage as evidence for more targeted treatment. As a result, MM diagnosis systems are desperately needed to improve clinical diagnosis efficiency.
Previously, many researchers have also focused their attention on the analysis of BM. In general, there are two types of methods: manual feature-based methods and deep learning methods. For manual feature-based methods, Saeedizadeh et al [15], extracted features from the nucleus and cytoplasm and then used a support vector machine (SVM) to classify malignant PCs. Benazzouz et al [16], used texture features and shape features to distinguish malignant PCs. Gupta et al. [17] developed an algorithm to segment PCs based on a color model-driven probabilistic multiphase level set. For these traditional methods, most of them only focus on the research of morphology classification, it is necessary to design complex feature extraction operators, which severely limits the automated myeloma analysis process. For deep learning methods. Tehsin et al. [18] used convolution to extract features and then used SVM to classify myeloma. Kumar et al. [19] detected white blood cancer from BM by using deep learning technology. Gehlot et al. [20] proposed a CNN framework with projection loss for MM diagnosis. Matek et al. [21] adopted deep neural networks to achieve the classification of various BM cells on a larger image. Although the above methods have achieved good results in BM cell classification, they are modeled as classification tasks, where manual identification and isolation of images with single cells from the whole slide are necessary, making the diagnoses extremely labor- and time-consuming. Consequently, the above myeloma analysis methods are not fully automated and are limited by the accuracy of single-cell extraction, which does not meet clinical needs.
In cell microscopic images, there might be some unideal situations that can seriously affect the performance of the detection, such as the nonuniformity of the cell staining, uneven distribution of the cells of different sizes in the slide, overlapping of the cells, half-cut cells at the edge of the slide. It is challenging to extract cells using unified segmentation algorithms [22]. Therefore, to overcome the limitations of existing methods, we introduced the notion of object detection into the task of myeloma analysis for the first time and proposed a novel end-to-end framework for fully automatic MM diagnosis on the whole slide. We transform BM cells classification challenges into multi-object detection problems. Therefore, we directly handle the whole slide and make the diagnoses completely automatic, which significantly relieves the pressure on manual cell detection and hence increases the efficiency of the diagnoses. In this work, we not only detected monoclonal PCs, but also other five different lineages of BM cells from the slide, including erythroid, granulocytic, lymphoid, monocytic cells, and others, so as to further assist the diagnosis of MM. Furthermore, we also used data enhancement technology to improve the problem of limited cell detection performance caused by uneven staining of microscopic images. In this study, we have obtained excellent cell detection performance on clinical test images. Our findings provide a novel approach for future automated cytological analysis work.
2. Methods
2.1 Sample processing and data preparation
This research collected the BM morphological images from 40 patients diagnosed with MM at the Department of Hematology, Zhongnan Hospital of Wuhan University between January 2021 to December 2021. Inclusion criteria were the diagnosis of multiple myeloma, the degree of active myelodysplasia or above, and the proportion of plasma cells in bone marrow ≥10%. The cohort included 23 males and 17 females, the age range of included patients was 38 to 85 years, with a median of 61.5 years. This study was approved by the Medical Ethics Committee of Zhongnan Hospital of Wuhan University (No. 2019065). All microscopic images taken from the BM of MM patients had given written informed consent for the use of clinical data according to the principles of the Declaration of Helsinki.
To obtain BM cells images for network training, as shown in Fig. 1, the following three steps were used: sample processing, image scanning, and image annotation. The first step was sample processing as shown in Fig. 1 we performed a BM aspiration on the patient to extract <0.2ml of BM fluid to make 6-8 BM smears and then added 2ml of Wright-Giemsa stain solution to cover the smear once it was dried. Next, we add 2-3ml PH6.4-6.8 phosphate buffer saline (PBS) buffer and mixed very well, then co-stained for 15-20 minutes at room temperature, and finally rinsed for drying. The second step was scanning imaging. We chose the proper target field of view and inserted the dried stained BM smear on the stage of the Olympus BX53 upright biological official microscope. Next, we scanned with a 100 ×oil immersion objective field of view (NA:1.25 WD 0.15spring) to enhance the magnification of the target field of view cells to 10 × 100, and then gathered images with a microscope. In this paper, we acquired 400 clear images of BM with a dimension of 2376 × 1824 pixels, with the physical size of the camera pixels being 10 × 10 um, totaling roughly 7000 cells. As shown in Fig. 2(A), We provide various instances of MM microscopic images and some distinct each lineage of cells corresponding to a label of 1-6 of single-cell images. According to clinical requirements, we divided all the cells into six categories: monoclonal PCs, erythroid, granulocytic, lymphoid, monocytic, and others. To facilitate network training, we classified various contaminants and smeared cells into other categories
Fig. 2. Each lineage cell corresponds to a label of 1-6 of cell image distribution. (A) Multiple myeloma microscopic imaging results. (B) Images distribution of different cell lineages.
The last step was image labeling. The initial cell image annotation was manually labeled by cytology specialists from the Department of Hematology, Zhongnan Hospital of Wuhan University, after obtaining high-quality and effective numerous BM microscopic images. Following labeling by the expert, student volunteers use the python LabelMe software to perform cell object detection and labeling, and each lineage cell corresponds to a label of 1-6. The correctness of the final image annotation was repeatedly checked by cytology specialists. In this study, as shown in Fig. 2(B), the training, validation, and testing sets were randomly divided into three groups in the ratio of 3:1:1. The validation set was used to prevent the model from overfitting and to finetune the hyperparameters. The test set was not involved in training and was only used to evaluate the accuracy and reliability of our method. Since there were fewer effective clinical images, thus, we used data augmentation technology to enrich the dataset and improved the robustness of the model.
2.2 Network and training
We present a novel automated analysis method for MM diagnosis, which transforms BM cells classification challenges into multi-object detection problems. In clinical applications, high-precision and efficient cytological analysis are usually required. Therefore, we use the current mainstream YOLOV5 network as the MM analysis framework [23], which is very lightweight and has high-precision detection. Moreover, it achieved the highest ranking ever in the object detection task, and it has fewer model parameters with 35.7M and lower floating-point operations per second (FLOPs) with 50.0G, which can perform image in highly accurate detection and high-throughput.
The proposed framework for MM diagnosis is shown in Fig. 3(A). Firstly, we performed data augmentation on the input images to improve the robustness of the model. Specifically, we used hue saturation value (HSV) color enhancement technology [24] to improve as much as possible the limited cell detection performance caused by uneven image staining, and simply apply common random rotation, translation, and zooming of images to further enhance the richness of the training set, Fig. 3(B) and (C) showed that the bounding boxes had various scales and there might be fewer cells on certain images, so we used novel image enhancement techniques including mix-up, mosaic and copy-paste [25] to heighten the performance of tinny object detection and improve the problem of category imbalance. Secondly, we used YOLOV5 as the BM cells detection network, which mainly included YOLO Backbone, YOLO Head, and detection module. YOLO backbone is composed of a five-layer feature pyramid network (FPN) [26], which can down-sampling the input image to obtain multiscale features. Specifically, the first layer comprises a convolutional layer [27], a batch normalization layer, and a sigmoid weighted liner unit (CBS) [28]. Each layer from the second layer to the fifth layer is composed of CBS and cross-stage partial (CSP) modules [29], CSP is an efficient feature fusion structure [29, 30] that can improve gradient transmission efficiency and support feature reuse, thereby further improving feature representation capabilities. In addition, we applied a spatial feature pyramid fusion (SPPF) [31] layer directly behind the YOLO backbone, which uses three max-pooling layers with different kernels to extract multi-scale features, contributing to solving the alignment problem of anchor and feature map. Moreover, SPPF can increase the receiving range of trunk features and focus on context features, thus enhancing the model's ability to detect various scale cells. Thirdly, YOLO Head combines low-level location information with high-level semantic information on a corresponding scale FPN map, enhancing semantic representation and positioning capabilities on multiscale. And the CSP module without shortcuts can reduce the amount of calculation while ensuring detection accuracy and speeding up network inference by integrating the gradient change into the feature map. Finally, since the low-level object location is accurate and high-level feature semantic information is more plentiful. To this end, in this paper, we performed detection on P2, P3, and P4 level heads, each of which generates a feature map by 1 × 1 convolution. Anchor boxes [32] are then applied to the output feature map, and a final predictions vector is generated with cell classification probability, objectness confidence score, and bounding boxes. Ultimately, we used non-maximum suppression (NMS) [33] to screen the bounding boxes to obtain optimal detection results.
Fig. 3. The proposed framework for multiple myeloma analysis and bounding boxes analysis. (A) YOLOV5 network in BM cells detection of MM patients. (B) and (C) are the center (x, y) coordinates and the length and width distribution of the object detection bounding boxes on the training set, respectively. (D) are the data enhancement results of the training set.
In the training phase, we employed binary cross-entropy (BCE) loss to calculate the class classification score and use CIoU loss for bounding box localization [34]. BCE loss is an objectness score to supervise objectness score, which is used to identify whether is there an object or not. To improve detection performance, YOLOV5 introduces the anchor mechanism, which provides prior knowledge about object bounding box size. The k-means algorithm was used to calculate adaptive anchor boxes in this experiment. In the experiment, batch size was set to 4, the input image was resized to 1280 × 1280 pixels with maintaining the aspect ratio, and the padding method was used to fill the original image with the black edge adaptively. The entire network was trained with stochastic gradient descent (SGD) [35] for 3000 epochs with an initial learning rate of 0.01, weight decay was 0.0005, and momentum is 0.937. And then using OneCycleLR [36] learning and warmup strategy to optimize the training process. We saved the best model weights when the validation average accuracy score was the highest. For the inference stage, we used the well-trained framework to detect the BM cells in the whole microscopic image. All the experiments were conducted in Pytorch [37] framework.
2.3 Evaluation metrics
To quantitatively analyze the cells detection performance, we used the Jaccard index, namely intersection over union (IoU) [38]to measure the accuracy of the object detection bounding box. IoU can be described as:
Precision represents false detection, that is, the misclassification of other kinds of cells as certain kinds of cells. The higher the value, the less the false detection rate. Conversely, Recall indicates missed detection, that is, a certain lineage of the cell is predicted as another lineage of the cell. The higher the value, the less the missing detection rate, F1 scores mean the harmonic mean of the Precision and Recall. Specifically, we considered a bounding box IoU score greater than 0.5 as true positives (TP), and an IOU score less than 0.5 as false positives (FP). False-negative (FN) represents a certain kind of cell detected as another cell or background. mAP denotes the average area of all classes under the PR curve, P represents Precision, R represents Recall, and m represents cell classes.
3. Experimental results
3.1 Performance
To detect the effective performance of the proposed method. As shown in Fig. 4, we performed diverse cells detection. Different lineages of cells detected are framed by the corresponding rectangle, surrounded with prediction score and confidence on the upper left side. By doing so, we found that this method effectively detected all cells, even the occluded cells were also well detected, thereby indicating that our newly developed method has a higher recall and detection performance. Furthermore, we also found that although some cell microscopic images were stained unevenly and certain kinds of cell targets were small, however, this novel approach proved effective to detect them. Importantly, the multinucleated monoclonal PCs were also correctly detected. The above analysis shows that our algorithm has good robustness and tiny object detection performance, as well as feature discrimination capability, and it also verifies the effectiveness of data enhancement.
Fig. 4. Examples of cell detection results and corresponding attention maps. (A) and (B) showing the analysis of network detection results on various cytological images. The detection results and attention masks are shown in the top row and bottom row, respectively. Among them, different lineages of cells are marked with different colored rectangles. 1-6 indicates monoclonal PCs, erythroid, granulocytic, lymphoid, monocytic cells, and others, respectively. The attention masks are generated by applying an attention map to corresponding images, the red area indicates the relevant potential features learned by the network. The green “${\color{green}{\star}}$” indicates misclassification.
In addition, to enhance the interpretability of our method, and to determine which areas of the input image were the focus of the neural network, as shown in Fig. 4, we visualized the attention feature map of the middle convolution layer. By doing so, we found that cell regions containing relevant features were colored red, suggesting that our model can focus on cell features while ignoring background features in microscopic images. From the perspective of cell morphology, lymphocytes are typically small cells with a round nucleus and little cytoplasm. However, monoclonal PCs usually have regular cell bodies, the shape of nuclei is generally round, and sometimes the nucleoli may appear dual or multinucleated. These cell morphological characteristics may increase the confidence and make the network pay attention to the relevant features of single-cell, resulting in obtaining high detection accuracy.
3.2 Quantitative analysis
As shown in Table 1, we exhibited the precision, recall, and F1 scores of the cell detection on the validation and testing set. Except for lymphoid and monocytic, all lineages of cells on the testing set achieved strongly detection scores consistent with the validation set across all the metrics. Meanwhile, this novel approach also performed well on the testing set for lymphoid. Although our method proved very effective for the detection of diverse cells, but couldn’t achieve a high precision rate for monocytes. This might be due to the diversity of monocytes, which could be easily confused with other cell lineages, including neutrophilic metamyelocytes, large lymphocyte cells, and monoclonal PCs, causing these cells to be classified as monocytes by mistake. However, it's worth mentioning that monocytes have a high recall rate, indicating that the present cells detection method has high detection performance and has a low misdetection rate.
Table 1. The class-wise precision, recall, and F1 scores of the neural network detection.
To gain further insight into the effectiveness and reliability of our newly developed method in MM cell detection, as shown in Fig. 5(A), we drew a radar map to highlight the difference in cell detection accuracy between using data augmentation techniques and not using them. Intuitively, we found that the cell detection performance was significantly improved by using data augmentation. Specifically, when only simple image rotation, translation, and zooming are used, then the cell detection effect will be relatively poor, among them, the monoclonal PCs have an accuracy rate of 84.34%, erythroid cells have an accuracy rate of 77.63%, the accuracy of granulocytic cells reach 70.68%, the accuracy rate is only 38.18% for the lymphoid cells, 34.83% for the monocytic cells, and 81.65% for the others. However, when using data enhancement, the corresponding kinds of cell detection accuracy rates are significantly enhanced by 8.11%, 13.6%, 20.99%, 43.64%, 51.73%, and 8.26%, respectively.
Fig. 5. Analysis of quantitative results of BM cells detection on the testing set. (A) The accuracy rate of cell detection with and without data augmentation technique. (B) PR curves without data augmentation (C) PR curves with data augmentation (D) confusion matrix as a manifestation of the cell detection results on the testing sets.
In addition, we also drew several cell lineages of precision-recall (PR) curves to further verify the effectiveness of the data augmentation technology. Figure 5(B) exhibited the PR curve without data enhancement technology. It showed that the overall trend of the curve appeared to be rather chaotic, which indicates that the detection accuracy of different lineages of BM cells is uneven and the cell detection effect is weak. All classes’ mean average precision (mAP) value was only 0.6625 at the IoU threshold of 0.5. On the contrary, except for monocytes, the general PR curve trend was relatively smooth when data augmentation was used; the possible explanation is that the proportion of monocytic cells is relatively low, which results in steepness near the middle of the curve. Moreover, as shown in Fig. 5(C)our algorithm achieved a good mAP score of 0.9173, to be specific, the average precision (AP) values of monoclonal PCs, erythroid, granulocytic, and lymphoid cells were all close to greater than 0.9 under different confidence thresholds, and the monocytic and other cells were also close to 0.85 or higher. The research confirmed the usefulness of the data augmentation methodology again, which improved the model's robustness and improved our method's capacity to detect cells. The results have revealed that the present method has a high cell detection precision and low false alarm. Conclusively, to analyze the BM cells detection results better intuitively, we also quantified the ability of our method to detect six lineages of BM cells, a confusion matrix is shown in Fig. 5(D), We found that monoclonal PCs achieved an accuracy rate of 92.45%, indicating that our newly developed method is expected to perform an intelligent diagnosis of MM. Moreover, other lineages of BM cells have also achieved an excellent accuracy rate. Specifically, the accuracy rate of erythroid cells reached 91.23%. The accuracy rate for the granulocytic cells achieved 91.67%, 81.82% for lymphoid cells, 86.21% for monocytic cells, and 89.91% for other cells. Together, these findings confirm that our newly developed method has achieved better detection performance in the microscopic image cell detection of BM cells. Finally, in terms of computational efficiency, our method only takes 12.5ms to detect an RGB microscopic image with 1280 × 1280 pixels, allowing it to achieve high-throughput detection and exceed clinical requirements.
Finally, to further demonstrate the advantages of our method, Table 2 presents a quantitative assessment of the recently relevant BM cell analysis methods. Obviously, we can see that the proposed method achieves better performance in terms of most quantitative metrics. Specifically, our method is oriented toward the automatic diagnosis of MM with a F1 of 93.6% for Monoclonal PCs. However, Gehlot et. al. [21] and Yu et. al. [22] mainly focus on BM cell classification and did not address malignant cell detection in MM. In addition, Gehlot et. al. performs single-cell image classification, where manual identification and isolation of images with single cells from the whole slide are necessary, making the diagnoses extremely labor- and time-consuming. Although Yu et. al. proposed a segmentation method for the analysis of whole-slide BM images, it is limited by the complex smear background and is difficult to apply to universal cell detection scenarios. Furthermore, although Matek et. al. [21] proposed a unified CNN framework for the diagnosis of MM, it can only achieve binary classification, i.e. monoclonal plasma cells and others. While our network can accurately detect as many as six lineages of cells from the whole slide, making it more applicable in clinical settings. The above evidence suggests that our approach provides a promising candidate for multiple myeloma diagnosis by significantly improving both the efficiency and the accuracy of exciting methods.
Table 2. Quantitative comparison with the recently relevant BM cells analysis methods. a
4. Conclusion and discussion
The microscopy of cell images for the diagnosis of MM is an indispensable method. Existing methods mainly focus on binary classification, that is, they mainly detect monoclonal PCs and others. However, if only two lineages of cells are detected, it is difficult to identify non-myeloma patients since PCs are far less common in normal patients. As we know, BM cells are composed of erythroid granulocytic, lymphoid monocytic cells, and others. In addition, Myeloma patients have a variety of clinical manifestations, with different symptoms leading to changes in the proportion of cells. Therefore, the detection of common cell lineages has an auxiliary role in the accurate diagnosis of MM, which is conducive to the discovery of suppressed cell lineage as evidence for more targeted treatment. However, there are challenges due to individual differences, staining differences, cell morphology diversity, and restricted expert knowledge. In terms of cell morphology, BM cells usually show a diversity of cell size and irregular morphology. For example, granulocytes are divided into myeloblasts, immature granulocyte, mature granulocyte, and monoclonal PCs also contain different cytoplasm and inclusions. The clinical diagnostic performance was severely hampered by the morphological variations of these cells. Therefore, we model multiple myeloma diagnoses as an objective-detection task. Therefore, we directly handle the whole slide and make the diagnoses completely automatic, which significantly relieves the pressure on the manual cell detection and hence increases the efficiency of the diagnoses. In practice, some unideal situations can dramatically affect the performance of the detection, such as the nonuniformity of the cell staining, uneven distribution of the cells of different sizes in the slide, overlapping of the cells, and half-cut cells at the edge of the slide. In our network, we design data preprocess methods to address the above problems to make sure the target cells can be correctly detected.
In general, the cell detection results are encouraging. In most cases, we achieved excellent MM diagnostic performance, and the accuracy of monoclonal PCs reached 92.45%. The average cell detection accuracy rate reached 90.8% and mAP with 91.17%, demonstrating the high accuracy of the proposed method. Specifically, the erythroid, granulocytic, lymphoid, monocytic cells, and others achieved a high accuracy rate of 91.23%. 91.67%, 81.82%, 86.21% and 89.91%, respectively (Fig. 5(D)). Compared with the recently reported classification of BM cells, our proposed approach has achieved a high detection accuracy rate. Moreover, we model multiple myeloma diagnoses as an objective-detection task. Moreover, our method only takes 12.5ms to detect a whole BM image, allowing it to achieve high-throughput detection and exceed clinical requirements. We firmly believe that this cell detection method is expected to become a clinical application, which can provide cytologists with auxiliary diagnosis, so as to save labor costs, and also provide a new idea for analogous cytological analysis applications. In the future, we will provide our model to the hospital for testing, aiming to help our research for clinical use as soon as possible, and we are also exploring more possibilities to build an ideal approach that may provide greater benefits to MM patients. Since all the multiple myeloma cell images in this paper were collected by clinicians, the sufficient sample available is limited by historical data.
Funding
National Natural Science Foundation of China (61905182, 62075200); Science Fund for Distinguished Young Scholars of Hubei Province (2021CFA042); Wuhan Research Program of Application Foundation and Advanced Technology (2020020601012237); Natural Science Foundation of Jiangsu Province (BK20221257); Fundamental Research Funds for the Central Universities (2042022kf1224); 2020 Medical Science and Technology Innovation Platform Support Project of Zhongnan Hospital of Wuhan University (lcyf202010); Translational Medicine and·Multidisciplinarv Research·Project·of·Zhongnan·Hospital of Wuhan University (ZNJC202217, ZNJC202232); Discipline Cultivation Project of Zhongnan Hospital of Wuhan University (ZNXKPY2022017); The Key Research and Development Program of Hubei province (2020BAB005); JSPS Core-to-Core Program.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. B. Barlogie, J. Shaughnessy, G. Tricot, J. Jacobson, M. Zangari, E. Anaissie, R. Walker, and J. J. B. Crowley, “Treatment of multiple myeloma,” Blood 103(1), 20–32 (2004). [CrossRef]
2. S. V. Rajkumar and S. Kumar, “Multiple myeloma current treatment algorithms,” Blood Cancer J. 10(9), 94 (2020). [CrossRef]
3. S. V. Rajkumar, “Treatment of multiple myeloma,” Nat. Rev. Clin. Oncol. 8(8), 479–491 (2011). [CrossRef]
4. W. Liu, J. Liu, Y. Song, X. Wang, M. Zhou, L. Wang, J. Ma, J. Zhu, and Union for China Leukemia Investigators of the Chinese Society of Clinical Oncology, Union for China Lymphoma Investigators of the Chinese Society of Clinical Oncology, “Mortality of lymphoma and myeloma in China, 2004–2017: an observational study,” J. Hematol. Oncol. 12(1), 22 (2019). [CrossRef]
5. S. V. Rajkumar, M. A. Dimopoulos, A. Palumbo, J. Blade, G. Merlini, M.-V. Mateos, S. Kumar, J. Hillengass, E. Kastritis, and P. J. T. L. O. Richardson, “International myeloma working group updated criteria for the diagnosis of multiple myeloma,” Lancet Oncol. 15(12), e538–e548 (2014). [CrossRef]
6. A. Lakshman, S. V. Rajkumar, F. K. Buadi, M. Binder, M. A. Gertz, M. Q. Lacy, A. Dispenzieri, D. Dingli, A. L. Fonder, and S. R. Hayman, “Risk stratification of smoldering multiple myeloma incorporating revised IMWG diagnostic criteria,” Blood Cancer J. 8(6), 59 (2018). [CrossRef]
7. E. Kastritis, E. Terpos, L. Moulopoulos, M. Spyropoulou-Vlachou, N. Kanellias, E. Eleftherakis-Papaiakovou, M. Gkotzamanidou, M. Migkou, M. Gavriatopoulou, and M. J. L. Roussou, “Extensive bone marrow infiltration and abnormal free light chain ratio identifies patients with asymptomatic myeloma at high risk for progression to symptomatic disease,” Leukemia 27(4), 947–953 (2013). [CrossRef]
8. S. V. Rajkumar, D. Larson, and R. A. . Kyle, “Diagnosis of smoldering multiple myeloma,” N. Engl. J. Med. 365(5), 474–475 (2011). [CrossRef]
9. A. S. Lagoo and N. S. Rosenthal, “Bone marrow at initial diagnosis: clinical associations and approach,” J. Hematop. 9(1), 9–18 (2016). [CrossRef]
10. J. N. Kather, A. T. Pearson, N. Halama, D. Jäger, J. Krause, S. H. Loosen, A. Marx, P. Boor, F. Tacke, and U. P. J. N. M. Neumann, “Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer,” Nat. Medicine 25(7), 1054–1056 (2019). [CrossRef]
11. A. J. Ohsaka, “Artificial intelligence (AI) and hematological diseases: establishment of a peripheral blood convolutional neural network (CNN)-based digital morphology analysis system,” Rinsho Ketsueki 61(5), 564–569 (2020). [CrossRef]
12. X. Bossuyt, M. Delforge, M. Reynders, D. Dillaerts, B. Sprangers, K. Fostier, K. Poesen, and M. J. L. Vercammen, “Diagnostic thresholds for free light chains in multiple myeloma depend on the assay used,” Leukemia 32(8), 1815–1818 (2018). [CrossRef]
13. S. V. Rajkumar, “Multiple myeloma: Every year a new standard?” Hematol. Cncol. 37, 62–65 (2019). [CrossRef]
14. P. Musto, M. Engelhardt, and J. J. H. Caers, “2021 European Myeloma Network review and consensus statement on smoldering multiple myeloma: how to distinguish (and manage) Dr. Jekyll and Mr. Hyde,” Haematologica 106(11), 2799–2812 (2021). [CrossRef]
15. Z. Saeedizadeh, A. Mehri Dehnavi, A. Talebi, H. Rabbani, O. Sarrafzadeh, and A. J. J. O. M. Vard, “Automatic recognition of myeloma cells in microscopic images using bottleneck algorithm, modified watershed and SVM classifier,” J. Micro. 261(1), 46–56 (2016). [CrossRef]
16. M. Benazzouz, I. Baghli, M. L. Benomar, and M. A. Chikh, “Automated identification of plasma cell in bone marrow images,” in Proceedings of Conference on Mathematics and Information Technology (ICMIT) (IEEE, 2017), pp. 8–12.
17. A. Gupta, P. Mallick, O. Sharma, R. Gupta, and R. J. P. O. Duggal, “PCSeg: Color model driven probabilistic multiphase level set based tool for plasma cell segmentation in multiple myeloma,” PLoS One 13(12), e0207908 (2018). [CrossRef]
18. S. Tehsin, S. Zameer, and S. Saif, “Myeloma cell detection in bone marrow aspiration using microscopic images,” in Proceedings of 2019 11th International Conference on Knowledge and Smart Technology (KST) (IEEE, 2019), pp. 57–61.
19. D. Kumar, N. Jain, A. Khurana, S. Mittal, S. C. Satapathy, R. Senkerik, and J. D. Hemanth, “Automatic detection of white blood cancer from bone marrow microscopic images using convolutional neural networks,” IEEE Access 8(12), 142521–142531 (2020). [CrossRef]
20. S. Gehlot, A. Gupta, and R. J. Gupta, “A CNN-based unified framework utilizing projection loss in unison with label noise handling for multiple Myeloma cancer diagnosis,” Med. Image. Anal. 72, 102099 (2021). [CrossRef]
21. C. Matek, S. Krappe, C. Münzenmayer, T. Haferlach, and C. J. Marr, “Highly accurate differentiation of bone marrow cell morphologies using deep neural networks on a large image data set,” Blood 138(20), 1917–1927 (2021). [CrossRef]
22. T. C. Yu, W. C. Chou, C. Y. Yeh, C. K. Yang, S. C. Huang, F. M. Tien, and S. C. Chou, “Automatic bone marrow cell identification and classification by deep neural network,” Blood 134(Supplement_1), 2084 (2019). [CrossRef]
23. Y. Fang, X. Guo, K. Chen, Z. Zhou, and Q. J. B. Ye, “Accurate and Automated Detection of Surface Knots on Sawn Timbers Using YOLO-V5 Model,” BioResources 16(3), 5390–5406 (2021). [CrossRef]
24. F. S. Hassan and A. J. Gutub, “Improving data hiding within colour images using hue component of HSV colour space,” CAAI Trans. on Intel Tech. 7(1), 56–68 (2022). [CrossRef]
25. I. Pacal and D. J. C. I. B. Karaboga, and Medicine, “A Robust Real-Time Deep Learning Based Automatic Polyp Detection System,” Comput. Biol. Med. . 134, 104519 (2021). [CrossRef]
26. T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (IEEE, 2017), pp. 2117–2125.
27. S. Albawi, T. A. Mohammed, and S. Al-Zawi, “Understanding of a convolutional neural network,” in Proceedings of Conference on Engineering and Technology (ET) (IEEE, 2017), pp. 1–6.
28. S. Elfwing, E. Uchibe, and K. J. Doya, “Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,” Neural Networks 107, 3–11 (2018). [CrossRef]
29. C.-Y. Wang, H.-Y. M. Liao, Y.-H. Wu, P.-Y. Chen, J.-W. Hsieh, and I.-H. Yeh, “CSPNet: A new backbone that can enhance learning capability of CNN,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPR) (IEEE, 2020), pp. 390–391.
30. K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (IEEE, 2016), pp. 770–778.
31. K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling in deep convolutional networks for visual recognition,” IEEE Trans. Pattern. Anal. Mach. Intel.37(9), 1904–1916 (2015). [CrossRef]
32. Y. Zhong, J. Peng, J. Wang, and L. Zhang, “Anchor box optimization for object detection,” in Proceedings of the IEEE Winter Conference on Applications of Computer Vision (ACV) (IEEE, 2020) pp.1286–1294.
33. R. Girshick, “Fast R-CNN,” in Proceedings of the IEEE Conference on Computer Vision (CVPR) (IEEE, 2015), pp. 1440–1448.
34. Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren, “Distance-IoU loss: Faster and better learning for bounding box regression,” in Proceedings of the Conference on Association for the Advancement of Artificial Intelligence (AAAI) (2020), 34(7), 12993–13000.
35. L. Bottou, “Stochastic gradient descent tricks,” in Neural Networks: Tricks of the Trade, G. Montavon, G. B. Orr, and K. R. Müller, eds. (Academic, 2012), pp. 421–436.
36. L. N. Smith and N. Topin, “Super-convergence: Very fast training of neural networks using large learning rates,” in Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications (International Society for Optics and Photonics, 2019), pp. 1100612.
37. A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, and L. J. Antiga, “Pytorch: An imperative style, high-performance deep learning library,” Adv. Neural Inf. Process Syst. 32, 8026–8037 (2019).
38. A. A. Taha and A. J. Hanbury, “Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool,” BMC Med. Imaging 15(1), 29 (2015). [CrossRef]
