Virtual University Journals
Search
Search
Automated Blood Cancer Classification with Lightweight CNNs Using Transfer Learning and Improved Data Augmentation
Dilawar Hussain, Maria Bibi, Asif Ali Durrani, Muhammad Adeel Ajmal Khan
Abstract:
Accurate diagnosis of blood cancer from microscopic blood smear images remain a careful and time-consuming task that depends on expert review. This study presents an auto- mated screening system based on lightweight convolutional neural networks that classify five blood cell types related to leukemia staging. Transfer learning with MobileNetV2 and EfficientNetB0 is used on a dataset of 5,000 high resolution images at 1024 by 1024 pixels from the Kaggle blood cell collection. The images are preprocessed by resizing them to 224 by 224 pixels and by applying contrast enhancement, followed by extensive data augmentation with geometric transforms, photometric changes and added noise. On the five-class validation set the tuned MobileNetV2 model reaches 94.42% accuracy, with overall precision of 95.8%, recall of 95.6% and F1 score of 95.6%, and a validation loss of 0.1701. The model converges in 10 epochs with a total training time of 343.26 minutes. These results indicate that lightweight CNN models can provide fast and accurate screening support in clinical settings and can help address the need for reliable automated diagnostic tools in hematology.
Keywords:
Blood cancer classification, Convolutional neural networks, Transfer learning, MobileNetV2, EfficientNetB0, Data augmentation
Full Text:
PDF