Deep-learning based labeling methods have gained unprecedented popularity in different computer vision and medical image segmentation tasks. However, to the best of our knowledge, these have not been used for cervical tumor segmentation. More importantly, while the majority of innovative deeplearning works using convolutional neural networks (CNNs) focus on developing more sophisticated and robust architectures (e.g., ResNet, U-Net, GANs), there is very limited work on how to aggregate different CNN architectures to improve their relational learning at multiple levels of CNN-to-CNN interactions. To address this gap, we introduce a Dynamic Multi-Scale CNN Forest (C K+1 DMF), which aims to address three major issues in medical image labeling and ensemble CNN learning: (1) heterogeneous distribution of MRI training patches, (2) a bidirectional flow of information between two consecutive CNNs as opposed to cascading CNNs-where information passes in a directional way from current to the next CNN in the cascade, and (3) multiscale anatomical variability across patients. To solve the first issue, we group training samples into clusters, then design a forest with (+ 1) trees: a principal tree of CNNs trained using all data samples and subordinate trees, each trained using a cluster of samples. As for the second and third issues, we design each dynamic multiscale tree (DMT) in the forest such that each node in the tree nests a CNN architecture. Two successive CNN nodes in the tree pass bidirectional contextual maps to progressively improve the learning of their relational non-linear mapping. Besides, as we traverse a path from the root node to a leaf node in the tree, the architecture of each CNN node becomes shallower to take in smaller training patches. Our C K+1 DMF significantly (p<0.05) outperformed several conventional and ensemble CNN architectures, including conventional CNN (improvement by 10.3%) and CNN-based DMT (improvement by 5%).

No abstract

The segmentation and classification of Lymph Nodes (LNs) is a fundamental but challenging step in the analysis of medical images of cervical cancer. Both tasks can leverage morphological features such as size, shape, contour, and heterogeneous appearance. However, these features might vary with the progressive state of LNs. Hence, accurate detection of LNs boundary is an essential step sing to classify LN as suspect (malignant) and non-suspect (benign). However, manual delineation of LNs might produce classification errors due to the inter and intra-observer variability. Semi-automatic and automatic LNs segmentation methods are greatly desired as they would help improve patient diagnosis and treatment processes. Currently, Magnetic Resonance Imaging (MRI) is widely used to diagnose cervical cancer and LN involvement. Diffusion Weighted (DW)-MRI exhibits metastatic LN as bright regions. This paper presents a semi-automatic segmentation and classification method of LNs. Specifically, we propose a novel approach which leverages (1) the complementarity of structural and diffusion MR images through a fusion step and (2) morphological features of the segmented metastatic LNs for classification. The contribution of our proposed algorithm is threefold. First, we fuse the axial T2-Weighted (T2-w) anatomical image and the DW image. Second, we detect LNs using region-growing method in order to compute the final classification. Third, segmentation results are then used to classify LNs based on a gray level dependency matrix technique which extracts LN features. We evaluated our method using 10 MR images T2-w and DW with 47 metastatic LNs. We obtained an average accuracy of 70.21% for cervical cancer nodule classification.

Dictionary Learning (DL) has gained large popularity in solving different computer vision and medical image problems. However, to the best of our knowledge, it has not been used for cervical tumor staging. More importantly, there have been very limited works on how to aggregate different interactions across data views using DL. As a contribution, we propose a novel cross-view self-similarity low rank shared dictionary learning-based (CVSS-LRSDL) framework, which introduces three major contributions in medical image-based cervical cancer staging: (1) leveraging the complementary of axial and sagittal T2w-magnetic resonance (MR) views for cervical cancer diagnosis, (2) introducing self-similarity (SS) patches for DL training, which explore the unidirectional interaction from a source view to a target one, and (3) extracting features that are shared across tumor grades and grade-specific features using the CVSS-LRSDL learning approach. For the first and second contributions, given an input patch in the source view (axial T2w-MR images), we generate its SS patches within a fixed neighborhood in the target view (sagittal T2w-MR images). Specifically, we produce a unidirectional patch-wise SS from a source to a target view, based on mutual and additional information between both views. As for the third contribution, we represent each individual subject using the weighted distance matrix between views, which is used to train our DL-based classifier to output the label for a new testing subject. Overall, our framework outperformed several DL based multi-label classification methods trained using: (i) patch intensities, (ii) SS single-view patches, and (iii) weighted-SS single-view patches. We evaluated our CVSS-LRSDL framework using 15 T2w-MRI sequences with axial and sagittal views. Our CVSS-LRSDL significantly (p<0.05) outperformed several comparison methods and obtained an average accuracy of 81.73% for cervical cancer staging. INDEX TERMS Shared dictionary learning, low-rank models, self-similarity, cross-view, cervical cancer stage.