医学知识嵌入的因果框架在急性白血病骨髓形态学AI辅助诊断中发挥关键作用

doi:10.16150/j.1671-2870.2026.01.007

摘要/Abstract

摘要：

目的：构建医学知识嵌入的因果人工智能（artificial intelligence，AI）框架模型，基于医学诊断逻辑进行急性白血病骨髓形态学AI辅助诊断。方法：本研究构建了医学知识嵌入的、分段式因果AI诊断框架模型。基于上海瑞金医院2020年至2023年间收治的93例病例（69例骨髓及外周血正常的其他疾病患者；24例血液疾病患者，其中14例急性白血病）骨髓及外周血涂片的20 089张显微图像数据集，以7∶3的比例训练和测试骨髓增生状态评估模块（模块1）、骨髓细胞种类识别模块（模块2）、外周血 AKP 组织化学染色积分评估模块（模块3）、POX 组织化学染色（模块4）及 CE组织化学染色（模块5）特征提取模块，用以提取多种关键骨髓形态学特征，包括骨髓增生状态、骨髓细胞种类识别、骨髓及外周血不同细胞组织化学染色程度等，并在2023年另10例真实病例的2 750张图中进行全流程验证。立足于血液系统疾病骨髓外周血形态诊断实践，对上述特定AI识别模块，根据医学诊断逻辑分步组合，整合多维度骨髓细胞形态学特征，进行两阶段序贯诊断预测（第一阶段：整合模块1、2、3；第二阶段：整合模块4、5）。第一阶段为正常、异常骨髓象病例鉴别，并同步识别慢性粒细胞白血病；第二阶段为急性白血病诊断及类型鉴别。同时建立端到端AI诊断框架模型作为对照，在真实病例中验证医学知识嵌入的因果AI诊断框架模型效能。结果：医学知识嵌入的因果AI诊断框架模型，在真实病例诊断测试中，准确率均高于端对端AI框架诊断模型（第一阶段准确率90.00%比70.00%，第二阶段准确率83.33%比66.67%）。结论：医学专业知识与诊断逻辑在血液系统疾病骨髓细胞形态学AI辅助诊断中发挥关键作用，应用医学知识嵌入的因果AI诊断框架模型，提升了AI辅助诊断急性白血病的准确率与可解释性，为血液系统疾病智能诊断提供了可推广的新范式。

关键词: 骨髓检查, 急性白血病, 形态学诊断, 深度学习, 因果AI学习框架

Abstract:

Objective To construct a causal artificial intelligence (AI) framework model embedded with medical knowledge for AI-assisted diagnosis of bone marrow morphology in acute leukemia based on medical diagnostic logic. Methods A stage-by-stage causal AI diagnostic framework guided by medical knowledge was constructed using a dataset of 20 089 microscopic images of bone marrow and peripheral blood smears of 93 cases(69 patients with other diseases exhibiting normal bone marrow and peripheral blood; 24 patients with hematological diseases, including 14 cases of acute leukemia ) from Ruijin Hospital (2020-2023). Multiple specialized modules including bone marrow hyperplasia assessment module(M1), bone marrow cell classification module(M2), AKP score calculation module(M3), POX-stained leukaemic cell feature extraction module(M4), and CE -stained leukaemic cell feature extraction module(M5), were trained and tested in a 7∶3 ratio to extract key morphological features, including bone marrow hyperplasia status, the constituent ratio of bone marrow cells, and cytochemical staining patterns in both bone marrow and peripheral blood. Based on morphological characteristics of the peripheral blood and bone marrow of hematological diseases, the aforementioned specific AI recognition models were combined step by step according to medical diagnostic logic, integrating multi-dimensional morphological characteristics of bone marrow cells, and a two-stage process sequential diagnosis and prediction was carried(stage 1: integrating M1, M2, and M3;stage2: integrating modules 4 and 5). The first stage involved distinguishing between normal and abnormal bone marrow cases, while simultaneously identifying chronic myeloid leukemia. The second stage focused on the diagnosis of acute leukemia and the differentiation of its subtypes. An end-to-end AI diagnostic framework model was also established as a control and compared with the medical knowledge-embedded causal AI diagnostic framework model in real-case diagnostic testing. Results The medical knowledge-embedded causal AI diagnostic framework model achieved higher accuracy than the end-to-end AI framework model in real-case diagnostic testing at both stages (Stage 1 accuracy: 90.00% vs. 70.00%; Stage 2 accuracy: 83.33% vs. 66.67%). Conclusions Medical expertise and diagnostic logic play a crucial role in AI-assisted diagnosis of bone marrow cell morphology for hematological diseases. The application of the medical knowledge-embedded causal AI diagnostic framework model improves the accuracy and interpretability of AI-assisted diagnosis, providing a novel and generalizable paradigm for intelligent diagnosis of hematological diseases.

Key words: Bone marrow examination, Acute leukemia, Morphological diagnosis, Deep learning, Causal AI learning framework

中图分类号:

李妙慧, 顾峻豪, 黄婷, 乔宇, 陈冰. 医学知识嵌入的因果框架在急性白血病骨髓形态学AI辅助诊断中发挥关键作用[J]. 诊断学理论与实践, 2026, 25(01): 44-52.

LI Miaohui, GU Junhao, HUANG Ting, QIAO Yu, CHEN Bing. Key role of causal framework embedded with medical knowledge in AI-assisted diagnosis of bone marrow morphology for acute leukemia[J]. Journal of Diagnostics Concepts & Practice, 2026, 25(01): 44-52.

图/表 6

图1

表1

表2

表3

表4

表5

参考文献 27

[1]	沈悌, 赵永强. 血液病诊断及疗效标准(第4版)[M]. 北京: 科学出版社,2018:87-113.
	SHEN T, ZHAO Y Q. Criteria for diagnosis and efficacy of hematological diseases (4th ed)[M]. Beijing: Science Press,2018:87-113.
[2]	熊树民, 余润泉. 临床血液细胞学图谱与应用[M]. 上海: 上海交通大学出版社,2009:15-19.
	XIONG S M, YU R Q. Atlas and application of clinical hematology[M]. Shanghai: Shanghai Jiao Tong University Press,2009:15-19.
[3]	CAO K, XIA Y, YAO J, et al. Large-scale pancreatic cancer detection via non-contrast CT and deep learning[J]. Nat Med, 2023, 29(12):3033-3043. doi: 10.1038/s41591-023-02640-w
[4]	AHN B, MOON D, KIM H S, et al. Histopathologic image-based deep learning classifier for predicting platinum-based treatment responses in high-grade serous ovarian cancer[J]. Nat Commun, 2024,15:4253.
[5]	ZHU L, SHI H, WEI H, et al. An accurate prediction of the origin for bone metastatic cancer using deep learning on digital pathological images[J]. eBioMedicine, 2023,87:104426.
[6]	CHANDRADEVAN R, ALJUDI A A, DRUMHELLER B R, et al. Machine-based detection and classification for bone marrow aspirate differential counts: Initial development focusing on nonneoplastic cells[J]. Lab Investig, 2020, 100(1):98-109. doi: 10.1038/s41374-019-0325-7 URL
[7]	YU T C, CHOU W C, YEH C Y, et al. Automatic bone marrow cell identification and classification by deep neural network[J]. Blood, 2019, 134(Suppl 1):2084. doi: 10.1182/blood-2019-125322 URL
[8]	MATEK C, KRAPPE S, MÜNZENMAYER C, et al. Highly accurate differentiation of bone marrow cell morphologies using deep neural networks on a large image data set[J]. Blood, 2021, 138(20):1917-1927. doi: 10.1182/blood.2020010568 pmid: 34792573
[9]	WANG C W, HUANG S C, LEE Y C, et al. Deep learning for bone marrow cell detection and classification on whole-slide images[J]. Med Image Anal, 2022,75:102270.
[10]	WANG W, LUO M, GUO P, et al. Artificial intelligence-assisted diagnosis of hematologic diseases based on bone marrow smears using deep neural networks[J]. Comput Meth Programs Biomed, 2023,231:107343.
[11]	WANG C, WEI X L, LI C X, et al. Efficient and highly accurate diagnosis of malignant hematological diseases based on whole-slide images using deep learning[J]. Front Oncol, 2022,12:879308.
[12]	ABHISHEK A, JHA R K, SINHA R, et al. Automated detection and classification of leukemia on a subject-independent test dataset using deep transfer learning supported by Grad-CAM visualization[J]. Biomed Signal Process Control, 2023,83:104722.
[13]	LITJENS G, KOOI T, BEJNORDI B E, et al. A survey on deep learning in medical image analysis[J]. Med Image Anal, 2017,42:60-88.
[14]	LIU M, LEE C W, SUN X, et al. Learning Causal Alignment for Reliable Disease Diagnosis[C]. arXiv, 2025.
[15]	YALCIN C, ABRAMOVA V, TERCEÑO M, et al. Hematoma expansion prediction in intracerebral hemorrhage patients by using synthesized CT images in an end-to-end deep learning framework[J]. Comput Med Imag Graph, 2024,117:102430.
[16]	ESTEVA A, KUPREL B, NOVOA R A, et al. Dermatologist-level classification of skin cancer with deep neural networks[J]. Nature, 2017, 542(7639):115-118. doi: 10.1038/nature21056
[17]	RICHENS J G, LEE C M, JOHRI S. Improving the accuracy of medical diagnosis with causal machine learning[J]. Nat Commun, 2020,11:3923.
[18]	PATIL R S, SZOLOVITS P, SCHWARTZ W B. Causal understanding of patient illness in medical diagnosis[M]. New York,Springer New York,1985:272-292.
[19]	SEVER C, ABBOTT C L, DE BACA M E, et al. Bone marrow synoptic reporting for hematologic neoplasms: Guideline from the college of American pathologists pathology and laboratory quality center[J]. Arch Pathol Lab Med, 2016, 140(9):932-949. doi: 10.5858/arpa.2015-0450-SA pmid: 26905483
[20]	SWERDLOW S H, CAMPO E, Harris N L, et al. WHO classification of tumours of haematopoietic and lymphoid tissues[M]. France: International Agency for Research on Cancer (IARC),2017:156-161.
[21]	LI N, FAN L, XU H, et al. An AI-aided diagnostic framework for hematologic neoplasms based on morphologic features and medical expertisep[J]. Lab Invest, 2023, 103(4):100055. doi: 10.1016/j.labinv.2022.100055 URL
[22]	CAI Z, VASCONCELOS N. Cascade R-CNN: High quality object detection and instance segmentation[J]. IEEE Trans Pattern Anal Mach Intell, 2021, 43(5):1483-1498. doi: 10.1109/TPAMI.2019.2956516 URL
[23]	LIU Z, LIN Y, CAO Y, et al. Swin transformer: Hierarchical vision transformer using shifted windows[C]. 2021 IEEE/CVF International Conference on Computer Vision (ICCV),2021:9992-10002.
[24]	王霄霞, 夏薇, 龚道元. 临床骨髓细胞检验形态学[M]. 北京: 人民卫生出版社,2019:3-90.
	WANG X X, XIA W, GONG D Y. Clinical bone marrow cell morphology[M]. Beijing: People's Medical Publishing House,2019:3-90.
[25]	SIDHOM J W, SIDDARTHAN I J, LAI B S, et al. Deep learning for diagnosis of acute promyelocytic leukemia via recognition of genomically imprinted morphologic features[J]. NPJ Precis Onc, 2021,5:38.
[26]	PAESSLER M E, HELFRICH M, WERTHEIM G B W. Cytochemical staining[M]. New York: Springer New York,2017:19-32.
[27]	KENNETH K, MARSHALL A L, JOSEF T P, et al. Williams hematology[M]. 9th ed. New York: McGraw-Hill Education, 2017.

Modules	Training set（images）	Testing set（images）	Total （images）
Bone Marrow Hyperplasia Assessment Module	140	84^a	224
Bone Marrow Cell Classification Module	9 881	4 234	14 115
AKP Score Calculation Module	1 409	1 251^b	2 660
POX-stained Leukaemic Cell Feature Extraction Module	1 480	634	2 114
CE-stained Leukaemic Cell Feature Extraction Module	683	293	976

Modules	Recall	Precision
Bone marrow cell classification module	84.36%	88.52%
POX-stained leukaemic cell feature extraction module	86.48%	90.89%
CE-stained leukaemic cell feature extraction module	85.48%	84.86%

Items	Image counts of Wright’s- stainedbone marrow (40×Magnification)	Image counts of Wright’ s-stained bone marrow (100× Magnification)	Image counts of AKP- stained peripheral blood	Image counts of POX-stained bone marrow	Image counts of CE-stained bone marrow
Normal (n=2)	14	250	210	/	/
CML-CP (n=2）	14	209	206	/	/
ALL (n=2)	14	220	154	136	135
APL (n=2）	14	190	188	108	110
AML (n=2）	14	260	46	128	130
Total	70	1129	804	372	375

Diagnostic framework	Training accuracy	Testing accuracy
Causal diagnostic framework	97.16%	90.00%
End-to-End diagnostic framework	93.48%	70.00%

Diagnostic framework	Training accuracy	Testing accuracy
Causal diagnostic framework	95.20%	83.33%
End-to-End diagnostic framework	100.00%	66.67%