Accurate anatomical landmark correspondence is highly critical for medical image registration. Traditionally many of the previous works proposed a number of hand-crafted feature sets that can be used to perform correspondence. However these feature tend to be highly specialized in terms of application area, and cannot be always generalized well to other applications without significant modifications. There have been other works that perform automatic feature extraction, but their reliance on labelled data hinders their ability to perform in cases where there is none.
To this end Wu et al. propose an unsupervised feature learning method which does not require labelled data. Their approach aims to directly learn the basis filters that can effectively represent all observed image patches. The learnt basis filters are later regarded as general image features representing the morphological structure of the patch.
In order to learn the basis filters, the authors propose a two-layer convolutional neural network called the Independent Subspace Analysis (ISA) algorithm. As an extension of ICA, the responses are not required to be all to be mutually independent in ISA. Instead, these responses can be divided into several groups, each of which is called independent subspace. Then, the responses are dependent inside
each group, but dependencies among different groups are not allowed. Thereby, similar
features can be grouped into the same subspace to achieve invariance. To ensure the accurate correspondence detection, multi-scale image features are necessary to use. However, it also raised a problem of high-dimensionality in learning features from the large-scale image patches.
This is achieved by constructing a two-layer network for scaling up the ISA to the large-scale image patches. Specifically, the ISA is first trained in the first layer based on the image patches with smaller scale. After that, a sliding window (with the same scale in the first layer) convolves with each large-scale patch to get a sequence of overlapped small-scale patches. The
combined responses of these overlapped patches through the first layer ISA are whitened by PCA and then used as the input to the second layer that is further trained by another ISA. In this way, high-level
understanding of large-scale image patch can be perceived from the low-level image features detected by the basis filters in the first layer.
The authors compare their work with two other methods, and apply their method on IXI and ADNI datasets.