Tissue Region Segmentation in H&E-Stained and IHC-Stained Pathology Slides of Specimens from Different Origins

In the era of advancing digital pathology, integrating digital slides and deep learning-based decision support systems has become increasingly prevalent in clinical practice. The effectiveness of these systems heavily relies on tissue region segmentation, a process essential for slide scanning and success of deep learning (DL) models. While thresholding-based methods are fast, they usually do not accurately detect tissue regions, especially in diverse staining scenarios or debris-laden images. Hence, this study develops a fast, lightweight, accurate, and robust deep-learning framework for segmenting tissue regions in digital pathology slides of various tissue types with different staining. The model is trained and tested on datasets of seven tissue types collected from seven different institutes. Internal and external validation studies have shown promising results, with area under receiver operating characteristics curve values greater than 0.98. The model has a memory footprint of 19.8 MBs, and it takes approximately 22 seconds to segment out a digital pathology slide using the CPU. The model could be used in digital slide scanners to improve the scanning process and in the pre-processing stages of DL pipelines to prepare high-quality datasets.