In medical image analysis, segmentation of the region-of-interest is the crucial phase for proper diagnosis. However, this task is very challenging due to missing or diffuse organ/tissue boundaries. However, deep learning frameworks based on the U-Net backbone have gained immense popularity since those can deal with image inconsistencies. This study compared U-Net’s performance with a marker-controlled watershed algorithm, a traditional segmentation algorithm. We tried to find which method is more efficient for polyp and overlapping cell segmentation. These two segmentation tasks were chosen since they are the most challenging task in medical history. The study of cell morphology is of great importance since it helps to understand the structure and conditions of cells, which helps us understand our underlying health conditions. A necessary before understanding cell morphology is cell segmentation. Polyps are protrusions from the inner lining of the colon or rectum which turns into colorectal cancer if left untreated. Thus, accurate segmentation of the polyp region is essential to help the physician properly diagnose the disease. In this work, we have used the KVASIR dataset to report the performance of the U-Net and watershed algorithm for polyp segmentation. A private dataset cultured on HeLa cell line was used for cell segmentation. Watershed algorithm obtained higher accuracy of 91.34% over U-Net which obtained accuracy 80.46% for cell segmentation. On the contrary, U-Net performed better for polyp segmentation and obtained accuracy 91.29% compared to watershed with reported accuracy of 89.19%.
CITATION STYLE
Roy, K., Bhattacharjee, D., Khatun, M., & Dutta, A. (2023). Domain-Specific Cues for the Usability of Marker-Controlled Watershed Algorithm and U-Net for Medical Image Segmentation. In Lecture Notes in Computational Vision and Biomechanics (Vol. 37, pp. 409–418). Springer Science and Business Media B.V. https://doi.org/10.1007/978-981-19-0151-5_34
Mendeley helps you to discover research relevant for your work.