Nucleus segmentation from the histopathological images of liver cancer through an efficient deep learning framework

Segmenting nuclei in histopathological images presents challenges due to variable sizes and overlapping regions, compounded by inter-class heterogeneity in shape and function. To overcome these, we proposed a robust deep learning architecture integrated with a novel edge detection technique based on local standard deviation. This method efficiently identifies nuclei edges, even at Multiscale levels, by exploiting the relationship between local standard deviation and image features. Our CNN architecture consists of stable residual, bottleneck, and attention decoding blocks, enabling effective extraction of high-level semantics and precise image localization. Evaluation on H&E stained liver disease histopathological images and a multi-organ database showcases the effectiveness of our approach, comparable to modern deep neural networks. Three performance indicators are computed to statistically assess segmentation efficacy, consistently demonstrating the superiority of our method over others. For multi-organ dataset segmentation using proposed methods attains the Jaccard index (89.99%), Precision (91.5%), Recall (91.1%), F1-measure (91.27%), and Accuracy (95.42%). Similarly for liver dataset attains Jaccard index (65.72%), Precision (88.74%), Recall (89.53%), F1-measure (90.12%), and Accuracy (94.85%). By addressing the limitations of previous unsupervised methods, our approach offers a robust solution for nuclei segmentation across diverse datasets. The significance of our work lies in its potential to enhance automated nuclei recognition in medical diagnosis and research, facilitating more accurate analysis of histopathological images and ultimately contributing to improved patient care and understanding of disease pathology.