Lin, Qiao
(2023)
Incorporating fuzzy-based methods to deep learning models for semantic segmentation.
PhD thesis, University of Nottingham.
Abstract
This thesis focuses on improving the workflow of semantic segmentation through a combination of reducing model complexity, improving segmentation accuracy, and making semantic segmentation results more reliable and robust. Semantic segmentation refers to pixel-level classification, the objective of which is to classify each pixel of the input image into different categories. The process typically consists of three steps: model construction, training, and application. Thus, in this thesis, fuzzy-based techniques are utilized in the aforementioned three steps to improve semantic segmentation workflow .
The widely-used semantic segmentation models normally extract and aggregate spatial information and channel-wise features simultaneously. In order to achieve promising segmentation performance, it is required to involve numerous learnable parameters, which increase the model's complexity. Thus, decoupling the information fusion tasks is an important approach in the exploration of semantic segmentation models. Fuzzy integrals are effective for fusing information, and some special fuzzy integral operators (OWA) are free of parameters and easy to implement in deep-learning models. Therefore, a novel fuzzy integral module that includes an additional convolutional layer for feature map dimensionality reduction and an OWA layer for information fusion across feature channels is designed. The proposed fuzzy integral module can be flexibly integrated into existing semantic segmentation models, and then help reduce parameters and save memory.
Following the exploration of semantic segmentation models, the collected data is used to train the model. Note that the precise delineation of object boundaries is a key aspect of semantic segmentation. In order to make the segmentation model pay more attention to the boundary, a special boundary-wise loss function is desirable in the segmentation model training phase. Fuzzy rough sets are normally utilized to measure the relationship between two sets. Thus, in this thesis, to improve the boundary accuracy, fuzzy rough sets are leveraged to calculate a boundary-wise loss, which is the difference between the boundary sets of the predicted image and the ground truth image.
After completing the training process with the proposed novel loss, the next step for semantic segmentation is to apply the pre-trained segmentation model to segment new images. One challenge is that there are no ground truth images to quantify the segmentation quality in the real-world application of semantic segmentation models. Therefore, it is crucial to design a quality quantification algorithm to infer image-level segmentation performance and improve the credibility of semantic segmentation models. In this thesis, a novel quality quantification algorithm based on fuzzy uncertainty is proposed as part of the model inference process without accessing ground truth images.
Moreover, to further explore the practical application of the proposed quality quantification algorithm in clinical settings, this thesis goes beyond public datasets and delves into a real-world case study involving cardiac MRI segmentation. Additionally, as clinicians also provide the level of uncertainty to measure their confidence when annotating to generate ground truth images (human-based uncertainty), the correlation between human-based uncertainty and AI-based uncertainty (calculated by the proposed quality quantification algorithm) is deeply investigated.
Comprehensive experiments are conducted in this thesis to demonstrate that the integration of fuzzy-based technologies can enhance the efficiency, accuracy, and reliability of semantic segmentation models compared to those without such methods.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
