A New Approach to Modify Post Transfer Learning with MobileNetV2 Architecture to Classify Acute Lymphoblastic Leukemia

Aim of the paper is to classify leukemic images using transfer learning with the MobileNetV2 architecture. Our strong point is to add global average pooling before flattening. To address the scarcity of labeled medical datasets, we employ transfer learning by using a pre-trained MobileNetV2 model that has been previously trained on a large dataset. We optimized the model's performance for leukemia classification by fine-tuning it on a large dataset of leukemia images. We have added global average pooling after transfer learning is finished. It is conducted to reduce the transfer learning weights' spatial data dimension and extracts each feature map into more stable information. We employ 10-2 , 10 -3, and 10-4 learning rates and only six epochs to classify leukemia images. We evaluated our proposed model using Acute Lymphoblastic Leukemia Image Database 2 (ALL-IDB2) through 5-Fold Cross-Validation (80% for training sets; the rest is employed as validation). Our model achieved remarkable results, with a 98% classification maximum accuracy. The results show that our maximum accuracy outperformed Linear - Support Vector Machine (SVM-L), Polynomial - Support Vector Machine (SVM-P), Radial Basis Function - Support Vector Machine (SVM-RBF). But it is not better than the Support Vector Machine - Convolutional Neural Network (SVM-CNN) and Convolutional Neural Network - Pyramid Model (PM-CNN).