RCMNet: A deep learning model assists CAR-T therapy for leukemia.

Area of Science:

• Hematology
• Immunology
• Biomedical Engineering

Background:

• Acute leukemia presents a high mortality rate, with current treatments facing recurrence challenges.
• Chimeric antigen receptor-T (CAR-T) cell therapy offers a promising avenue for treating acute leukemia.
• Accurate identification of CAR-T cells is critical for effective therapy but is hindered by their similarity to other blood cells.

Purpose of the Study:

• To develop a reliable method for identifying CAR-T cells to support their therapeutic application in leukemia.
• To create and evaluate a novel deep learning model for CAR-T cell morphological identification.

Main Methods:

• A dataset of 500 CAR-T cell microscopy images was constructed.
• A novel integrated model, RCMNet (ResNet18 with Convolutional Block Attention Module and Multi-Head Self-Attention), combining CNN and Transformer architectures, was developed.
• Transfer learning was applied to adapt RCMNet for the CAR-T cell dataset.

Main Results:

• RCMNet achieved 99.63% top-1 accuracy on a public image classification dataset.
• On the CAR-T cell dataset, RCMNet achieved a maximum accuracy of 83.36% using transfer learning, outperforming other state-of-the-art models.
• The model demonstrated satisfactory performance in image classification tasks.

Conclusions:

• RCMNet shows significant potential for accurate CAR-T cell identification.
• The developed model can be translated for diagnostic applications in clinical settings.
• This work addresses a key challenge in CAR-T cell therapy, paving the way for improved leukemia treatments.