DSTL-NET: Multi-level Leukemia Stage Classification Via Dual-Stage Segmentation Based Transfer Learning Networks

Leukemia is a life-threatening haematological malignancy caused by the aberrant proliferation of abnormal white blood cells, where early and timely diagnosis is important in improving survival outcomes. However, traditional diagnostic methods such as manual blood smear analysis are prone to human error, require extensive expertise, time constraints and difficulty in capturing subtle morphological variations across different stages. To overcome these challenges, a novel DSTL-Net is proposed for efficient classification of leukemia using blood smear images from publicly available Kaggle dataset. The proposed model initially integrates with Fine-level Laplacian filter (FLLF) to remove the noisy artifacts for improving the image quality by preserving cell morphology. Dual-stage W-net is introduced with boundary refinement modules for accurate leukemic cell segmentation. The transfer learning models (ResNet, EfficientNet, CapsuleNet) are used to extract the fine hierarchical features from the segmented leukemic cells. The Boruta-Lasso (BoLo) algorithm for selecting the optimal feature subset from the retrieved features. Finally, the fully connected layer processed the selected features for classifying the leukemia classes into normal, benign, early, pre-leukemic and pro-leukemic cases. Experimental evaluation on Kaggle dataset demonstrates that the proposed DSTL-Net achieves 98.44% accuracy with MCC of 0.97, dice score of 0.975 and IoU of 0.940. In addition to this, BoLo improves the computational efficiency by removing redundancy and reducing training complexity. Moreover, the proposed DSTL-Net increases the overall accuracy of 3.58%, 3.10%, 2.16%, 1.22%, 1.35% and 0.42% better than SVM, NasNetLarge + VGG, YOLOv2 + DeepLabv3, DarkNet + ShuffleNet, UNet and FOADCNN-LDC respectively.