A fine-grained image classification algorithm based on self-supervised learning and multi-feature fusion of blood...

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Leukemia diagnosis relies on manual blood cell image analysis, which is subjective and time-consuming.
• Current methods face challenges in accuracy and efficiency, potentially leading to misdiagnosis.

Purpose of the Study:

• To develop an automated blood cell image classification method for accurate leukemia diagnosis.
• To improve upon existing diagnostic tools by leveraging advanced deep learning techniques.

Main Methods:

• Utilized Masked Autoencoders (MAE) for self-supervised pre-training on TMAMD and Red4 datasets.
• Employed an enhanced Vision Transformer with feature fusion from all encoder layers.
• Incorporated subcenter Arcface Loss with dynamic margins for improved feature representation.

Main Results:

• Achieved state-of-the-art classification accuracies: 93.51% on TMAMD and 81.41% on Red4.
• Demonstrated effective utilization of multi-level features, including color, texture, and semantics.
• Showcased enhanced fine-grained feature discrimination.

Conclusions:

• The proposed MAE-enhanced Vision Transformer method significantly improves blood cell image classification for leukemia.
• This approach offers a promising, objective, and efficient tool to aid physicians in clinical diagnosis.
• The findings represent a valuable reference for advancing automated medical diagnostics.