Biomedical Image Segmentation Using Deep Learning

EE 610: Image Processing, Prof. Amit Sethi

The central aim of the project was to perform nuclei segmentation for breast tumor cells in histopathology images. This was achieved through the utilization of a multi-organ transfer learning approach, specifically tailored to segment the nuclei of breast tumor cells. Coded a U-Net architecture tailored for binary semantic segmentation for distinguishing between nucleus and non-nucleus regions in stained tissue images by employing TensorFlow+Keras. The U-Net model was trained using Diceloss and the architecture was also modified the architecture to study the effects of changing the number of blocks and filters per layer.

In the next phase of the project Watershed Segmentation was applied to the probability map produced by U-Net for segmentation of individual nuclei and demonstrated its effectiveness on test images containing overlapping or touching nuclei.