Morphological diagnosis of hematologic malignancy using feature fusion-based deep convolutional neural network

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Leukemia, a cancer of white blood cells, poses a significant mortality risk, necessitating early detection of blast cells.
• Current diagnostic methods can be time-consuming, highlighting the need for advanced tools to improve efficiency and accuracy.

Purpose of the Study:

• To develop and evaluate a novel deep convolutional neural network (CNN) named 3SNet for the automated diagnosis of leukemia cells.
• To assess the efficacy of 3SNet in identifying blast cells using combined image features for improved diagnostic support.

Main Methods:

• A deep CNN model, 3SNet, incorporating depth-wise convolution blocks was designed to reduce computational load.
• The model utilizes three input types: grayscale images, Histogram of Oriented Gradients (HOG) for shape, and Local Binary Patterns (LBP) for texture analysis.
• The 3SNet model was trained and validated on the AML-Cytomorphology_LMU dataset.

Main Results:

• The model achieved a Mean Average Precision (MAP) of 84% for cell types with fewer than 100 images and 93.83% for those with over 100 images.
• The Area Under the Receiver Operating Characteristic (ROC) curve exceeded 98% for the tested cell types, indicating high diagnostic performance.
• These results demonstrate the model's capability in accurately classifying leukemia cells based on their morphological features.

Conclusions:

• The developed 3SNet model shows significant potential as an adjunct tool for leukemia cell diagnosis.
• Its high accuracy and efficiency suggest it can serve as a valuable second opinion for clinicians, potentially improving patient outcomes.
• Further integration into clinical workflows could enhance the early detection and management of leukemia.