EfficientNetV2-S Enhancement to Classify Acute Lymphoblastic Leukemia: Integrating Pre-Trained Models and Grid Search for Optimal Performance

Leukemia is one of the deadliest types of cancer. The hospitals have been working on several occasions to conduct early screenings of the preventable death. Unfortunately, leukemia detection is very expensive. This study tries to classify Leukemia images to find the best value of the hyperparameter using the Grid Search method processed using Pre-trained EfficientNetV2-S. We have modified the activation function ReLu6 (Variant of the Rectified Linear Unit has value less than 6) to SeLu6 (Variant of the Scaled Exponential Linear Unit has value less than 6) to help the network maintain stable statistical properties during training. Our proposed model has five major stages: Input layer, Stem Layer, Mobile Inverted Bottleneck Convolution (MBConv), Fused Mobile Inverted Bottleneck Convolution (Fused-MBConv) and Head layer. We use a total of five different Fused-MBConv layers. The depth-wise convolutional architecture and expansion operation are fused into one single unified step. The process can be applied to improve the general performance to be more reliable and precise in getting responses. Our proposed way adopts a comprehensive scaling approach to adjust the depth, width, and image resolution proportionally. Besides, MBConv and Fused-MBConv are applied to enhance the performance of the model. We utilize Grid Search to perform hyperparameter tuning and obtained α = 0.001, and E = 10 as the optimal hyperparameter for our proposed architecture model. Our proposed model has been tested on the C-NMC-2019 Leukemia image dataset. Experimental in the training process have achieved accuracies 98.89% to 99.80%. The validation results give an accuracy within the range of 96.65% to 98.31%, while the testing results produce accuracy within the range of 98.01% to 99.85%. The AUC values for all folds have constantly generated an area of 0.97. We compare our proposed results with other methods, and the comparison results have shown that our performance results are better than EfficienNetB0, CNN-based ECA Module, Vision Transformer, Majority Voting Technique, CNN Model-based Tversky Loss Function, and Lightweight EfficientNet-B3. This research results can be followed up as product innovation in medical fields.