View-collaborative fuzzy soft subspace clustering for automatic medical image segmentation

Abstract

With the rapid development of medical imaging methodologies, such as magnetic resonance (MR) and positron emission tomography (PET)/MR, various types of MR images, which are acquired using inconsistent MR pulse sequences on the same patient, have been applied in medical-image-based diagnoses. A feature map extracted from an MR image describes the patient’s condition from one perspective. By effectively using all feature maps from various MR images, it is possible to completely describe the intrinsic characteristics of the patient’s condition to facilitate a diagnosis. Facing such a scenario, classic machine learning algorithms typically stack these feature maps for unified processing and do not explore the importance of each feature within a single feature map or the relationships among feature maps. To address these challenges, both multiview and subspace learning scenarios are considered in this study, and the multiview collaboration-based fuzzy soft subspace clustering (MVC-FSSC) algorithm is proposed. The MVC-FSSC algorithm not only strives to exploit the agreement of decisions across all views via collaborative learning but also strives to utilize the soft subspace-based weighting mechanism to automatically evaluate the contribution of each dimensional feature to each estimated cluster within a single view. Our experimental results indicate that the proposed MVC-FSSC algorithm can effectively explore the collaborative relations among all views and the importance of features in their respective views. Additionally, our MVC-FSSC method has substantial advantages over traditional clustering algorithms in MR image segmentation. Applying the MVC-FSSC algorithm to five patients’ MR images, the average mean absolute prediction deviation (MAPD) is 98.62 ± 8.34, which is significantly better than the score of 131.90 ± 16.03 that was obtained using the collaborative fuzzy k-means (CO-FKM) algorithm and the score of 128.87 ± 11.32 that was obtained using the quadratic weights and Gini-Simpson diversity-based fuzzy clustering (QWGSD-FC) algorithm.