A multi-scale feature fusion of deep learning network for classifying acute leukemia genetic subtypes from blood smear images

Le résumé fourni par la source

Acute leukemia (AL) is a common type of hematological malignancy. Accurate classification of its genetic subtypes is crucial for achieving precise clinical diagnosis and treatment. However, some genetic subtypes lack distinct morphological features and can easily be confused with other AL types. Additionally, genetic subtypes often relies on expensive genetic testing equipment. These factors limit the ability to achieve precise diagnosis of AL in clinical settings. To tackle this clinical challenge, we developed the EMAResDrop model by enhancing pre-trained ResNet50 with DropBlock and Efficient Multi-Scale Attention (EMA) techniques to improve the classification accuracy of genetic subtypes of AL. Our model was evaluated on two AL datasets comprising blood smear images from 469 cases of acute lymphoblastic leukemia (ALL) and 269 cases of acute myeloid leukemia (AML) patients. In the test set for ALL, the EMAResDrop model effectively distinguished between BCR::ABL1, TCF3::PBX1, KMT2A rearrangement, ETV6::RUNX1, and ALL-Others, achieving accuracies of 0.967 ± 0.005, 0.954 ± 0.006, 0.946 ± 0.005, 0.981 ± 0.006, and 0.929 ± 0.002, respectively. Additionally, the classification accuracies for different AML subtypes were 0.992 ± 0.013 for PML::RARA, 0.929 ± 0.004 for RUNX1::RUNX1T1, and 0.926 ± 0.005 for CBFB::MYH11. When compared with four benchmark models (VGG16, InceptionV3, DenseNet121, and ResNet50), EMAResDrop exhibited superior performance, achieving classification accuracies of 0.954 ± 0.015 and 0.949 ± 0.003 for the ALL and AML datasets, respectively. The external test results further confirmed the strong generalization ability of the model. Overall, this study enhanced the genotyping diagnosis of AL and has the potential to improve early diagnosis and treatment, potentially supporting more effective clinical management.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.