Xin Wang ¹ Gui R. Xie ^2* Wen Wang ^1* Jiao Y. Zhang ^3* Guang Y. Cong ^4* Han Xiang ^5* Yin Y. Zeng ^5* Lin X. Fang ^5* Xin S. Li ^1*

• ¹ Dongguan First Hospital affiliated to Guangdong Medical University, Dongguan, China
• ² The Centre for Medical Genetics and Key Laboratory of Maternal and Child Metabolic Genetics, Dongguan Maternal and Child Health Hospital, Dongguan, China
• ³ School of Basic Medicine, Guangdong Medical University, Dongguan, Guangdong Province, China
• ⁴ Clinical Laboratory, The First Affiliated Hospital of Guangdong Pharmaceutical University, Guangzhou, China
• ⁵ School of Biomedical Engineering, Guangzhou Medical University, Guangzhou, China

Chromosome classification plays a crucial role in identifying chromosomal abnormalities and is considered a challenging and significant task in karyotype analysis. Although deep learning techniques have shown promising results in chromosome classification tasks, their reliance on stacked convolutional blocks for extracting low-resolution high-level features may limit their ability to utilize high-resolution low-level semantic features for accurate characterization. To overcome this limitation and achieve precise chromosome classification, we propose a novel architecture called Multi-Scale Dual-Attention Network (MSDA-Net). MSDA-Net integrates a multi-scale feature extraction module and a multi-scale fusion module. The multi-scale feature extraction module includes several nested dense dual-attention modules, which not only capture feature information at different scales but also identify more discriminative features in terms of position and spatial dimensions. Subsequently, the multi-scale fusion module combines the acquired feature representations to predict chromosome image into one of the classes. In the chromosomal classification task, when evaluated on a large dataset of 24 types of images containing normal and abnormal single chromosomes that we created, our method demonstrates impressive accuracy, precision, recall, and F1-score of 99.03%, with good classification performance for all 24 types of chromosomes. Furthermore, extensive comparative experiments validate the superiority of our method over the current state-of-the-art convolutional neural network algorithms and can achieve the best classification results with fewer parameters and shorter training time. This study has significant clinical implications for the identification of diseases associated with abnormal chromosomes.