Cancer Cytoplasm Segmentation from Cell Image Based on Transfer Learning

Image segmentation is one of the most popular approaches used in digital pathology. Most existing deep learning-based image segmentation models extract nuclear from the background in cell images. The nuclear of cancer cell images are irregular in shape and size, with unclear borders. Detecting cancer regions from cell images for detailed analysis depending on the diagnosis's purpose, including other regions, is also necessary. Because of the characteristics of cancer cell images, conventional segmentation models trained on existing datasets targeted to extract cell regions only fail to detect cancer cytoplasm in cell images. Therefore, in this research, H&E stained cancer cell images are labeled manually, and segmentation is performed by transfer learning using a conventional method to segment cancer cytoplasm in cell images. Experimental results confirm that the IoU and dice coefficient are improved by 14% and 2%, respectively, compared to the method without transfer learning.