Kernels boost medical imaging
Osteoarthritis and breast cancer are diseases associated with high socioeconomic costs. Early detection could vastly improve patient prognosis and quality of life. However, medical imaging is still not powerful enough to detect such diseases in all patients due to problems with tissue heterogeneity and accuracy. The AKMI (Advanced kernel-methods for medical imaging) project addressed these issues. AKMI researchers comprehensively studied different feature learning architectures, including hierarchical kernel methods, restricted Boltzmann machines (RBMs) and convolutional neural networks (CNNs). In particular, their RBM solutions are already being applied in a myriad of pattern recognition tasks and are the building blocks of deep belief networks. Project members also applied and analysed kernel-based learning algorithms such as support vector machines (SVMs). They published an article that describes the risk of overfitting from improper model selection and also provide guidelines to avoid this. Another major achievement was the development of strategies for model comparison and selection that were published in the Journal of Machine Learning Research in 2015. Using linear classifiers like SVMs and the rho-margin loss function, they proved that minimising a basic margin-bound is NP-hard. They also used hashing algorithms to speed up kernel methods for the computation of set and bit-string similarities. AKMI successfully applied their machine learning algorithms based on CNN architecture for efficient 3D imaging of osteoarthritis and quantification of cartilage deterioration. This approach proved more accurate than MRI, the current state-of-the-art for the diagnosis of osteoarthritis. The application of CNN methods for breast cancer risk assessment also showed good performance. The AKMI study has proven that machine learning methods for pattern recognition and analysis can significantly improve medical image analysis. Implementation of such tools in clinical management of patients could lead to faster diagnosis, early treatment and better patient outcomes.
Keywords
Kernel, medical imaging, machine learning, osteoarthritis, breast cancer, AKMI,
