Diagnosing acute promyelocytic leukemia by using convolutional neural network.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate diagnosis of acute promyelocytic leukemia (APL) is critical for effective treatment.
• Traditional diagnostic methods for APL rely on expert interpretation of bone marrow smear images.
• Developing automated systems can enhance diagnostic efficiency and consistency.

Purpose of the Study:

• To evaluate the efficacy of instance segmentation using convolutional neural networks (CNNs) for diagnosing APL.
• To assess the performance of Mask R-CNN in identifying and classifying nucleated cells in bone marrow smears.
• To compare the diagnostic accuracy of models trained from scratch versus pre-trained and fine-tuned models.

Main Methods:

• Utilized a dataset of 13,504 bone marrow smear images.
• Employed the Mask R-CNN instance segmentation method for cell detection and classification.
• Trained models from scratch and fine-tuned a pre-trained Mask R-CNN model on the dataset.
• Applied data augmentation techniques to enhance model performance.
• Developed diagnostic systems based on trained models and the French-American-British (FAB) Classification.

Main Results:

• The best-performing model, an augmented pre-trained Mask R-CNN, achieved an average precision (AP) of 62.5% and an average recall (AR) of 84.1%.
• The pre-trained model demonstrated significantly higher average precision than the model trained from scratch (P < 0.05).
• Dataset augmentation further improved diagnostic accuracy (P < 0.03).

Conclusions:

• Instance segmentation with Mask R-CNN shows potential for accurate APL diagnosis in bone marrow smear images.
• A pre-trained approach combined with data augmentation significantly enhances diagnostic performance.
• Deep learning offers a promising avenue for improving APL detection systems.