Optimized Interactive Image Segmentation with An Interpretation Method
Z Chen1, X Qi2, S Tan1*, (1) Huazhong University of Science and Technology, Wuhan, China,(2) UCLA School of Medicine, Los Angeles, CA, USA
Presentations
SU-IePD-TRACK 1-2 (Sunday, 7/25/2021) 5:30 PM - 6:00 PM [Eastern Time (GMT-4)]
Purpose: A common challenge of deep-learning-based interactive segmentation methods is the need of retraining the original network after user’s clicks. The retraining process is computationally expensive, and the labels that the user provides can be easily down-weighted during retraining, leading to inaccurate segmentation results. In this study, we propose a novel interactive method, instead of performing optimization over the weights of a neural network, to optimize the activation in the intermediate layer. Meanwhile, we introduce the gradient-based network interpretation method to guide our optimization.
Methods: We use two types of operations to optimize the activation, namely Activation Replacement and Activation Reparameterization. We adopt the gradient-based interpretation method to find the spatial regions and feature channels in the activation that are associated with the target category. We replace the activation value in these spatial regions with the feature value corresponding to the category (called Activation Replacement). Besides, we introduce channel-wise scaling and bias to these associated feature channels as auxiliary parameters. The auxiliary parameters are optimized by the backpropagating refinement scheme (called Activation Reparameterization). Finally, we obtain the segmentation mask of the target object by updating the activation and performing the forward inference process again. The performance is evaluated over 120 MRI sequences for liver, kidney, and spleen segmentation.
Results: We use the deep object selection (DOS) algorithm as the baseline, which is a conventional deep-learning-based method. The performance of the interactive segmentation algorithms is evaluated by the Number of Clicks when the segmentation mask achieves 90% IOU over the ground truth (NoC@90). The average NoC@90 scores are 3.20 and 5.60 for our proposed method and the DOS respectively.
Conclusion: We propose an interactive segmentation algorithm by performing optimization over the activation in the intermediate layer. Experimental results suggest satisfactory performance achieved using our method, compared with the baseline.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by National Natural Science Foundation of China (NNSFC), under Grant Nos. 62071197 and 61672253.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email