1. A Novel Hybrid Convolutional and Network Encapsulation Approach in EfficientNetV2-S Architecture for Acute Lymphoblastic Leukemia Classification.
- Author
-
Muntasa, Arif, Sugiarti, Alisa, Wahyuningrum, Rima Tri, Husni, Ghufron, M. Ali, Hermawan, Almohamedh, Refan Mohamed, Motwakel, Abdelwahed, Asmara, Yuli Panca, Dewi, Deshinta Arrova, and Tuzzahra, Zabrina
- Abstract
Acute Lymphoblastic Leukemia (ALL) is one of the most common cancers among children under the age of 15. An increasing awareness to detect early leukemia can be avoided preventable deaths from leukemia. Accurate leukemia cells image classification plays a crucial role in reducing health risks by enabling early diagnosis, which helps in mitigating the severity of the disease through timely and targeted treatment The detection of Acute Lymphoblastic Leukemia incurs significant costs in terms of time. We have developed an alternative method, A Modified EfficientNetV2-S to detect Acute Lymphoblastic Leukemia quickly, accurately, and affordably. We proposed a new architecture called Hybrid Kernel on EfficientNetV2-S. Our architecture integrates three convolutions with different kernels: regular convolution, dilated convolution, and depth-wise convolution. Three convolutions work together in one layer to provide detailed information about the object. The Network Encapsulation method uses convolution results to find object features from an image based on different locations and directions and provide combined convoluted information. We update the weight to customize the Networks Encapsulation feature map results iteratively. We continuously repeat this operation until the updated weight difference falls below the specified threshold. The approach we suggested is unique because it uses several kernels for convolutions to help capture various feature variations. We have evaluated our new architecture using the C-NMC-2019 dataset at three learning and epoch pairs: {0.001;5}, {0.0001;10}, and {0.00001;15}. We used 12,528 images to assess the reliability of our proposed new architecture. The results showed that the best accuracy, precision, recall, and F1-Score were 97.44%, 97.35%, 98.88%, and 97.48%, respectively. Our proposed method performance outperformed VGG19, Xception, RestNet50, and EfficientNetB0. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF