Loading...

Instance Segementation in Medical Images Using Weak Annotation

Sadeghi, Mohammad Hossein | 2021

197 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 54729 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Behroozi, Hamid; Mohammadzadeh, Nargesol Hoda
  7. Abstract:
  8. Recent approaches in the field of semantic image segmentation rely on deep networks that are trained by pixel-level labels. This level of labeling requires a lot of time for the labeler person; because these networks require large training datasets to achieve optimal accuracy and the lack of data at the labeled pixel level causes a significant drop in their performance. In order to overcome this problem, weakly supervised segmentation approaches have been proposed. In these approaches, weaker labels such as image-level labels, bounding boxes, scribbles, etc. have been introduced to train the networks.In this thesis, a method for segmentation of kidney and kidney tumor in CT scan images based on training data tagged at the image level is introduced. To determine more precisely the slices of the CT scan image in which there is a kidney, video segmentation networks such as I3D have been applied along with the proposed cost function. Then, in order to determine the kidney pixels, a classification network is trained with CAM method and by using post-processing methods, we achieve a more accurate segmentation of kidneys. To optimize CAM performance, an adaptive thresholding method has been proposed. By finding kidney pixels in the relevant slices, an unsupervised method for detecting tumor pixels in slices labeled with tumors has been introduced, with the help of which we can appropriately obtain the extent of the kidney tumor. Finally, the bounding boxes detected as the tumor area by the unsupervised method, are used to train a detection network whose structure is based on transformer. With the help of the introduced structure, we are able to compare kidney areas and kidney tumors with comparable accuracy (Dice score equal to 0.861 in kidney segmentation and 0.643 in kidney tumor segmentation) with the best neural networks under full supervision (Dice score equal to 0.974 in kidney segmentation and 0.851 in kidney tumor segmentation
  9. Keywords:
  10. Transformers ; Semantic Segmentation ; Image Segmentation ; Tumor Segmentation ; Weakly Supervised Semantic Segmentation ; Medical Images ; Image-Level Label

 Digital Object List

 Bookmark

No TOC