Automated bone marrow cell classification through dual attention gates dense neural networks

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate bone marrow cell morphology analysis is crucial for diagnosing malignant hematological disorders.
• Current automatic classification models using convolutional neural networks show promise but lack clinical precision.
• Low accuracy limits the widespread clinical adoption of automated bone marrow cell classification.

Purpose of the Study:

• To develop a novel, efficient, and high-precision bone marrow cell classification model.
• To enhance the performance of automated bone marrow cell classification for improved diagnostic accuracy.

Main Methods:

• A Dual Attention Gates DenseNet (DAGDNet) model was developed by integrating a dual attention gates (DAGs) mechanism into DenseNet architecture.
• DAGs were employed to filter and emphasize position-related features within DenseNet, aiming to boost precision and recall.
• The model was trained and validated using a bone marrow cell morphology dataset from the First Affiliated Hospital of Chongqing Medical University, including leukemia samples, and a separate bone marrow cell classification dataset.

Main Results:

• The proposed DAGDNet demonstrated superior performance compared to established models like DenseNet and ResNeXt on a multi-center dataset.
• DAGDNet achieved a mean precision of 88.1% on the Munich Leukemia Laboratory dataset, establishing state-of-the-art performance.
• The model maintained high efficiency while achieving top-tier classification accuracy.

Conclusions:

• The DAGDNet model effectively enhances the accuracy of automatic bone marrow cell classification.
• This AI tool can serve as a valuable assisting diagnosis tool in clinical practice.
• DAGDNet offers an efficient solution for rapid inspection of numerous bone marrow cells, reducing the likelihood of diagnostic errors.