Optimizing deep learning for accurate blood cell classification: A study on stain normalization and fine-tuning techniques
BACKGROUND: Deep learning’s role in blood film screening is expanding, with recent advancements including algorithms for the automated detection of sickle cell anemia, malaria, and leukemia using smartphone images. OBJECTIVES: This study aims to build the artificial intelligence (AI) models and asse...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Wolters Kluwer Medknow Publications
2025-01-01
|
| Series: | Iraqi Journal of Hematology |
| Subjects: | |
| Online Access: | https://journals.lww.com/10.4103/ijh.ijh_110_24 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | BACKGROUND:
Deep learning’s role in blood film screening is expanding, with recent advancements including algorithms for the automated detection of sickle cell anemia, malaria, and leukemia using smartphone images.
OBJECTIVES:
This study aims to build the artificial intelligence (AI) models and assess their performance in classifying blood film cells as normal or abnormal.
MATERIALS AND METHODS:
The dataset included 171,374 images from 961 patients which were classified by experts. These images were resized, denoized, normalized, augmented, and classified into two categories, normal and abnormal cells. Two stain normalization techniques were used in this study; Reinhard and Mackenko techniques. The data were split into training and testing sets with a ratio of (8:2). The model was built through transfer learning by using the pretrained model Inception-Resnet v2 as a backbone. Three different fine-tuning techniques were tested in this study. The training was done using Python with Keras library on Google Colab for 10 epochs. The model was tested for accurately classifying individual blood cells whether normal or abnormal and evaluated using accuracy and area under receiver operator characteristic curve.
RESULTS:
The counts of the three most common cell types were as follows: Segmented neutrophils: 29,424; erythroblasts: 27,395; and lymphocytes: 26,242. The Reinhard stain normalization had better accuracy than Mackenko, the best AI model achieved the highest accuracy of 96.7%%, the area under the curve (AUC) of 99.87%, while the second technique achieved an accuracy of 91.46% and an AUC of 97.23% in classifying normal from abnormal cells.
CONCLUSION:
In conclusion, AI can effectively classify the blood cells as either normal or abnormal, yielding accurate results in a time-effective manner, especially with the use of transfer learning of pretrained models and fine-tuning. In this study, Inception-Resnet V2 showed good accuracy in differentiating normal from abnormal cells. |
|---|---|
| ISSN: | 2072-8069 2543-2702 |