Classification of normal, AML, and ALL bone marrow smears based on deep learning and hyperspectral microscopic imaging

Abstract

Bone marrow smear examination is critical in leukemia diagnosis, yet traditional manual methods face reliability and efficiency challenges. This study proposes a multimodal open-set recognition method that integrates deep learning with hyperspectral microscopy imaging. By analyzing the proportions of different cell categories, the method automatically classifies bone marrow smears into normal, acute myeloid leukemia (AML), and acute lymphoblastic leukemia (ALL). Bone marrow smear samples from 120 patients were collected and imaged using a hyperspectral microscopic imaging system. Individual cells were successfully segmented using U²-Net and image registration algorithms, resulting in spectral and image data for a total of 6606 cells. Based on these data, two one-dimensional convolutional neural networks (1D CNNs) architectures were designed for spectral data, two two-dimensional convolutional neural networks (2D CNNs) architectures were created for image data, and a dual-branch heterogeneous architecture consisting of parallel spectral and image branches was developed to simultaneously process both spectral and image data. Experimental results demonstrated that the optimal dual-branch heterogeneous architecture model based on fused spectral-image data achieved an accuracy of 91.57 % in open-set external validation, and attained 100 % accuracy in distinguishing normal, AML, and ALL in a subset of 19 bone marrow smear samples. Thus, by combining deep learning and hyperspectral microscopy imaging technology and introducing an open-set recognition method, this study achieved high-accuracy differentiation of bone marrow smears, significantly enhancing leukemia diagnostic efficiency and accuracy, and providing new insights and methods for leukemia diagnosis and classification.