Hybrid Techniques for the Segmentation and Classification of White Blood Cancer Based on the Fusion of Deep Network Features

Microscopic diagnosis of acute lymphoblastic leukemia (ALL) and multiple myeloma (MM) depends on expert review, and the public datasets available for automated analysis are small and imbalanced. This study presents a four-stage pipeline for nucleus segmentation and cell classification on the SN-AM and SegPC-2021 datasets. Images were enhanced with contrast-limited adaptive histogram equalization (CLAHE); a conditional generative adversarial network (pix2pix) generated 2500 synthetic image–mask pairs per dataset from new nucleus layouts; nuclei were segmented with UNet using five pre-trained CNN encoders and compared with YOLO26; and single cells were classified with XGBoost on fused, PCA-reduced deep features. Models were trained with and without synthetic data and evaluated on real test images only. The synthetic pairs improved UNet on SN-AM (Dice 0.934 to 0.951 for EfficientNetB7, six test images), whereas on SegPC-2021 their effect was within ±0.01 Dice; YOLO26 gave the best SegPC-2021 result (Dice 0.799). Fused features classified ALL and MM cells with a balanced accuracy of 0.978–0.984. However, background patches without leukocytes also identified the source dataset (ROC-AUC ≥ 0.996); in a same-source experiment, three fused feature sets still reached 0.993–0.996. Synthetic data helped most where annotations were scarce; larger same-source datasets are needed for confirmation.