A Robust Hybrid Preprocessing Pipeline for Overlapped Cell Segmentation in Peripheral Blood Smear Images

Abstract

Peripheral blood smear analysis is a vital diagnostic step for identifying hematological disorders such as Acute Lymphoblastic Leukemia (ALL). Overlapped cell regions present significant challenges in automated detection and segmentation, often reducing classification accuracy. This paper proposes a robust hybrid preprocessing pipeline integrating comparative denoising (Median, Gaussian, Wiener, bilateral, adaptive median) with advanced contrast enhancement methods (Histogram Equalization, CLAHE, and Contrast Stretching). Multiple hybrid filtering sequences are evaluated for optimal noise removal, quantified through PSNR, SNR, and MSE metrics. Segmentation is performed using both conventional approaches (thresholding, watershed, clustering) and deep learning-based methods (U-Net, Mask R-CNN) to handle cell overlap effectively. Experimental results on ALL image datasets demonstrate superior segmentation accuracy, structural detail preservation, and improved downstream classification performance. The proposed methodology enhances the reliability of automated blood smear analysis and offers a scalable solution for clinical laboratories seeking rapid and accurate leukemia detection.