GlandNet: Automated Gland Segmentation for Colorectal Cancer Diagnosis using UNet

Histopathological image study is vital for diagnosing diseases like cancer by the correct identification and analysis of cellular structures such as glands. Nevertheless, segmentation of these structures in particular for distorted or densely packed glands remains a huge challenge for most automation methods. Even though deep learning based algorithms look promising, they continue struggling with segmented gland morphologies, while traditional methods present challenges in adequately separating tightly clustered glands. This paper proposes GlandNet, a UNet based architecture that provides instance segmentation for better gland segmentation. By undertaking these tasks in conjunction, this approach enhances generalization, especially in difficult cases. In this work, GlandNet is evaluated using different metrices, establishing new benchmarks of state-of-the-art scores: 95%, 95%, and 71.35 pixels, respectively. Further comprehensive experiments on the benchmark GlaS dataset demonstrate that, particularly in delicately separating densely clustered glands, GlandNet indeed outperforms the existing state of affairs. This development increases the performance and efficiency of histopathological slide analysis, leading to better disease diagnostics and medical research.