A Semi-Supervised Learning Method Using Soft-Label for Cell Nuclei Segmentation On Immunohistochemistry Images
J Zhou1*, Z Yan2, J Polf3, H Zhang4, B Zhang5, M MacFarlane6, D Han7, M Zakhary8, A Gopal9, J Xu10, S Lee11, H Xu12, G Lasio13, S Chen14, (1) University of Maryland Shore Medical Center at Easton, Easton, MD, (2) Sense Brain, Princeton, NJ,(3) University of Maryland School of Medicine, Baltimore, MD, (4) University of Maryland School of Medicine, Baltimore, MD, (5) University of Maryland School of Medicine, Baltimore, MD, (6) University of Maryland School of Medicine, Baltimore, MD, (7) University of Maryland School of Medicine, Baltimore, MD, (8) University of Maryland School of Medicine, Baltimore, MD, (9) University of Maryland School of Medicine, Baltimore, MD, (10) University of Maryland School of Medicine, Baltimore, MD, (11) University of Maryland School of Medicine, Baltimore, MD, (12) University of Maryland School of Medicine, Baltimore, MD, (13) University of Maryland School of Medicine, Baltimore, MD, (14) University of Maryland School of Medicine, Baltimore, MD
Presentations
PO-GePV-M-291 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: We developed a semi-supervised learning method using soft-label for cell nuclei segmentation on Immunohistochemistry (IHC) images.
Methods: The proposed segmentation method was applied on IHC NuClick dataset 2020. This data set comprises of lymphocyte annotations within 871 IHC images and the ground truth labels. 2D images were extracted from the datasets and augmented 10 times and divided into three groups: 70% training images, 20% validation images, 10% test images. Within the training images, 50% images were used as labeled data and 50% images were used as unlabeled data. A renovated mean-teacher semi-supervised framework was developed which encourages the predictions to be consistent during various noises for the same input. The framework included two the identical architecture networks called a student model and a teacher model. The student model was regularized by a consistency soft loss with the teacher model and a supervised segmentation loss with the labels. After the weights of the student model were updated with gradient descent, the weights of the teacher model were updated and an exponential moving average (EMA) of the student weights.
Results: The quantitative results of our segmentation method achieved mean dice score of (0.65, 0.66), mean accuracy of (0.77, 0.78), mean recall of (0.64, 0.66), mean specificity of (0.83, 0.84), and mean precision of (0.64, 0.66) with 95% CI on the test datasets.
Conclusion: The qualitative and quantitative comparisons show that our proposed method can achieve higher segmentation accuracy with less variance on testing datasets. It will be useful in image analysis applications for cell nuclei pathology and radiotherapy assessment in IHC images.
Keywords
Segmentation, Image Analysis, Radioimmunotherapy
Taxonomy
IM/TH- Image Segmentation Techniques: Machine Learning
Contact Email