Weighted Robust PCA for Statistical Shape Modeling

Self-learning shape recognition in medical images

“…Experimental results show a boost of performance of WRPCA compared to RPCA refer to our previous conference publication [27]. However, several limitations exist that WRPCA is apparently restricted to the prior knowledge, i.e., the probability of each landmark of being an outlier.…”

“…In Chapter 3, an overview of the matrix factorization techniques addressing the problem of data contamination is provided, followed by the pipeline of statistical shape modeling using the proposed Kernelized RPCA. [18]; retinopathy [19] Abdomen liver [20]; kidney [21]; pancreas; spleen [22]; prostate [23,24] Skeletal knee cartilage [25]; vertebrae [26]; hip [26]; foot [27]; head and neck [28]; hand [29] Pathology Tumor brain tumor [30]; prostate cancer [31]; breast cancer [32,33]; skin lesion [34]; cervical cell […”

“…In [29], a shape and appearance model is derived for X-ray hand segmentation, where the segmentation is optimized by a single global criterion under the Maximum a posteriori inference. In our publications [27,95], robust SSMs are derived for MRI ankle bone segmentation which aims to cope with the data contamination and capture nonlinear patterns respectively.…”

“…The remaining of this chapter is organized as follows: first, a concise overview of matrix factorization techniques addressing the training data contamination is presented, followed by the introduction of statistical shape modeling using RPCA and our early contributions Weighted RPCA [27] as well as KPCA; subsequently, the proposed KRPCA is illustrated in detail; afterwards, experiments are conducted to validate the proposed method's ability of recovering outliers in training data, and evaluate the quality of model derived using competitive approaches as well as segmentation accuracy using these models. To be mentioned, evaluation results show that the model built with the proposed KRPCA has a higher quality than other state-of-the-art methods.…”

“…The idea is to replace the identified outliers with the mean value derived from the population, where the outliers are determined in terms of data frequency in distribution. Based on RPCA, a Weighted RPCA (WRPCA) is proposed in [27], the motivation behind is leveraging some prior knowledge to separate outliers and preserve underlying data in the constructed low-rank subspace.…”