Background & Summary

The evaluation of bone marrow smear cell morphology is the basis for the diagnosis of blood diseases1. It is particularly important in the case of MDS2,3, where pathological hematopoiesis causes cell morphology to be variable and complex4,5,6. Moreover, the test requires two pathologists to manually differentiate and count at least 500 nucleated cells on Wright-stained bone marrow smears7, which is labor-intensive and time-consuming8,9.

In the past decade automatic morphology analyzers have been used in laboratory work10,11,12. Although development in machine learning has successfully improved the accuracy and efficiency of the performance of these analyzers13, its improvement largely depends on the availability of large and high-quality datasets14,15,16. However, most studies have focused on peripheral blood cells, and the datasets of bone marrow cells are rare17,18,19,20. Given the variety of morphological features and numerousness of bone marrow cell types, some of which have low occurrence, it is hard to obtain a sufficient number of cells for effective algorithm training21,22. In the process of cell segmentation, bone marrow smears may be more difficult than peripheral blood smears owing to cell overlap23. Some studies have reported the complexity of unsatisfactory performance in real-world data using small datasets24. Therefore, the establishment of disease-specific bone marrow cell datasets will benefit the development of AI for cell morphology applications.

The dataset contains 25,009 bone marrow cell images from 125 patients with MDS. The image labels consisted of 32 categories, including normal cells and pathological hematopoietic cells, such as small megakaryocytes and micromegakaryocytes. Cell labels were reviewed by 3 experts. This dataset provides images under different shooting conditions, making it an ideal dataset for testing the robustness of machine learning models. It is particularly suitable for in-depth analysis of cell morphology evaluation in MDS and similar hematological diseases. The cell images are wider than actual cell margins by 10% to ensure cell integrity and to support random cropping in the data augmentation process before training. The publication of this dataset aims to address the limitations of existing datasets and provide more comprehensive data for the application of machine learning models in the evaluation of bone marrow cell morphology in hematological diseases, especially in the diagnosis of MDS.

Methods

The first-time bone marrow smears of 125 patients diagnosed with MDS from the First Affiliated Hospital of Zhejiang University Medical College were collected. (Institutional review board approval number No. 2024-0889) As the bone marrow smear does not contain identifiable personal information, it is exempted from obtaining written informed consent based on this situation. Among the 125 patients with MDS, 69 were male and 56 were female, aged 23–82 years, with a median age of 64 years. MDS was diagnosed based on the World Health Organization (WHO) 2022 classification criteria for hematopoietic and lymphoid neoplasms. Among them, 44 patients were MDS-LB subtype, 46 were MDS-IB1, 31 were MDS-IB2, 3 were MDS-h and 1 was MDS-f.

Sample preparation

Bone marrow smears were prepared manually by the wedge method. A BaSO BSZ-CM16-B automatic slide staining machine with BaSO Wright-Giemsa staining solution (solution a) and BaSO phosphate buffer (ph6.8) (solution b) was used to complete Wright-Giemsa staining of the smears.

Data acquisition

Slides were processed using MorphoInsight, a versatile and automated slide scanning analysis system developed by Kinsight, Co., Ltd., and successor of MorphoGo25,26,27. The system is equipped with slide loading/unloading mechanics, a moving gripper that carries the slide while scanning, and a high-speed camera. Its optical components include a 100X oil objective lens and an LED light source. For each slide, a whole-slide image (WSI) with a pixel distance of 75 nm was generated. Subsequently, the cell detection algorithm of the system obtained cell locations on the WSI as bounding boxes. Medical technologists manually selected the region of interest (ROI) for each slide. After converting the bounding boxes to squares according to longest side and expanding them by 10%, cells in the ROI were cut from the WSI. This step ensured the integrity of cell borders and morphological features within the cut images while minimum surrounding cells were included.

Additional images were collected during routine work using digital cameras (either SmartV550Df, SmartV650Df or SmartV650D) mounted on a microscope with 100X oil objective lens. The light source of the microscope is either a halogen lamp or an LED bulb. These images were processed with similar manner. Cells were located by algorithm, and their locations were confirmed by medical technologists. The bounding boxes were also converted to squares and expanded by 10% before the cells were cut from the images. The entire data-acquisition process is illustrated in Fig. 1.

Fig. 1
Fig. 1
Full size image

Data acquisition process diagram.

Data annotations

First, a medical technician with approximately eight years of work experience at hospital identified the subtypes of bone marrow nucleated cells. Subsequently, these annotations were reviewed by a medical expert with approximately 20 years of experience. In cases of inconsistency, the cells were rechecked by an expert with 25 years of experience. If no consensus was reached, the image was excluded from the dataset. If the staining quality was considered too poor or the cell segmentation was not suitable for identification, the image was excluded as well. Cell images were classified into 32 categories, including normal bone marrow cells and abnormal cells caused by pathological hematopoiesis in MDS, including blasts, dysplastic erythroblasts, dysplastic granulocytes, small megakaryocytes, and micromegakaryocytes (Fig. 2). These cells are commonly observed and are characteristic of MDS, representing the cell types that pathologists frequently utilize in their routine work for cell classification. Each MDS patient has different dysplastic cell manifestations, and not every type of dysplasia may be present in one patient28. Considering the relatively low occurrence of dysplastic cells, if further subdivided, the number of each cell type will be too small to meet the needs of AI training. And some pathological types are difficult to distinguish, but they can be distinguished from normal cells. Therefore, we consider classifying them by the 3 main myeloid series, grouping pathological anomalies of each series together. We expect classification systems trained with this dataset could inform the operator of abnormalities.

Fig. 2
Fig. 2
Full size image

Some annotated bone marrow cells in the database (a–e) blast cells, (f) ring-nucleus neutrophil, (g) binucleated granulocyte, (h) granulocyte with nuclear-cytoplasmic maturation asynchrony, (i,j) neutrophil with Pseudo-Pelger-Huët-like anomaly, (k) erythroblast with nuclear budding, (l) erythroblast with basophilic stippling, (m) trinucleated erythroblast, (n) erythroblast with Howell-Jolly bodies, (o) dysplastic erythroblast, (p–t) micromegakaryocytes, (u–y) small megakaryocytes.

Data Records

The root directory contains two folders. Cell images acquired through MorphoInsight are stored in ‘main’ folder, and additional cell images in ‘add’ folder. In each folder, images are grouped according to cell classification. Image names are formatted using cell classification and an index. Each folder contains an .csv file listing the image paths and corresponding information. For the .csv file in ‘main’ folder, cell classification and diameter are provided. And the .csv file in ‘add’ folder also contains camera resolution setting used to capture the original image.

Technical Validation

Images acquired through MorphoInsight were randomly split into training, validation, and test sets at ratios of 70%(15987 images), 10%(2352 images), and 20%(4503 images), respectively. Transfer learning using MorphoInsight’s cell classification model, which specializes in normal bone marrow cell classification as the base model, yielded an overall accuracy(number of cells whose model prediction was consistent with expert annotation, divided by total cell number) of 80.13%. During the training, small megakaryocytes, micromegakaryocytes, dysplastic granulocytes, and dysplastic erythroblasts were added as new types. The f1 scores (calculated as harmonic mean of recall and precision) of these four types of dysplastic cells were 50.85%, 64.94%, 31.25%, and 59.31% respectively. Their recall and precision rates are presented in Table 1. Some confusion between the dysplastic cells and their corresponding normal types was observed in the results, which could be explained by the morphological similarities between them. Figure 3. The base model cannot differentiate dysplastic granulocytes and erythroblasts from normal ones. And it predicted small and micro- megakaryocytes as mostly lymphocytes and smudge cells.

Table 1 The recall, precision, and F1 score of four special cell types of the model trained with this dataset.
Full size table
Fig. 3
Fig. 3
Full size image

Classification results of the trained model by means of the confusion matrix. Rows represent the annotated cell types and the columns the predicted type given by the classifier. Diagonal values (green cells) are the true positive number for each class. Out of diagonal values give the numbers of wrong predictions.

The test demonstrated that an acceptable classification performance of dysplastic cells can be achieved using our dataset. Better results can be expected with more cell images. In fact, the dataset offered here forms only a small portion of the data collected, as each WSI could contain thousands of cells. But the remaining parts have not been thoroughly annotated. It can be used for further learning and optimization of MorphoInsight’s cell classification model.

Usage Notes

The dataset consists of JPEG images and information files in CSV format. No special software is required to view and use this dataset.

As a result of their low occurrence, the number of some cell classes in this dataset is small and may be insufficient to achieve acceptable classification results. If desired, classes such as megaloblastic late erythroblast, histiocyte and dysplastic megakaryocyte could be excluded from training set. Classes similar in morphology and clinical significance could be merged. For example, blast NOC, myeloblast, monoblast and promonocyte could all be treated as blast.