Biological cell counting using deep learning

Abstract

Accurate biological cell counting plays a pivotal role in numerous biomedical applications, yet conventional manual and rule-based approaches struggle with dense, overlapping, and morphologically diverse cells. This thesis presents a hybrid deep learning framework for automated cell counting using both convolutional and transformer-based architectures. Initially, a Dual Cascaded Network (DCNet) is proposed, combining a VGG16-based encoder with a U-Net decoder to generate high-resolution density maps from microscopy images. To address limitations in crowded cells, a transformer-based alternative—Restormer—is employed, offering improved global context modeling through attention mechanisms and specialized components such as Multi-Dconv Head Transposed Attention and Gated Feed-Forward Networks. The study introduces Focal Inverse Distance Transform (FIDT) maps to enhance localization precision in dense cell environments. Additionally, a SALW strategy is integrated to dynamically balance learning difficulty across spatial regions. Evaluated on diverse datasets—including synthetic bacterial, bone marrow, and adipose tissue images—the proposed models demonstrate robust performance, achieving competitive accuracy across varying imaging conditions. This work highlights the effectiveness of hybrid architectures and attention-guided learning in advancing the state-of-the-art in cell counting.