Improved Anomaly Detection In Low-Resolution And Noisy Whole-Slide Images Using Transfer Learning

Abstract— Whole slide imaging (WSI) is one of the most recent technologies introduced in medical pathology practices. WSI images are created using a computerized system that scans, stitches and storages pathology specimen glass slides into digital images, which provide a multi-resolution pyramid construction of a huge gigabyte size due to the need for containing a high amount of tissue details. Therefore, digital WSI brings major challenges in data storage, image analysis and transmission (e. g. telepathology and interoperability). In this paper, we propose a computer-aided diagnosis (CAD) system to detect cancer anomalies in breast lymph node WSI images under low-resolution (LR) and noise conditions. In particular, we investigate a transfer-learning approach to find the scale mappings between WSI levels using partial least-square (PLS) regression. The learned scale mappings can be used to detect anomalies in LR images and hence reduce the computational cost of anomaly detection. Then we explore the effect of different levels of noise. We simulate different scenarios where WSI images are contaminated with Gaussian noise and several de-noising algorithms were applied, namely de-noising with PLS, Block Matching 3D (BM3D) and the combination of PLS and BM3D. We show that these de-noising algorithms can help reduce the noise severity on anomaly detection. For example, noisy images with 0. 8 noise standard deviation, these three algorithms improved the detection accuracy from 63. 50% to 93. 81%, 92. 73%, and 97. 51%, respectively. Our results lead to useful conclusions on how to handle whole slide images under scaling and noise conditions.

Keywords— Whole slide imaging, multi-resolution pyramid construction, partial least-square, computer-aided diagnosis, lymph node cancer.

Introduction

Whole slide imaging (WSI) is a novel emerging technology for high-resolution microscopic image analysis and telepathology. In this technology, which is also called virtual microscopy, images are viewed without the use of a microscope or slides. Virtual slides have a structure of multi-layers in multiple-resolutions at each image location, according to the highest resolution obtained to allow more detailed storage and display for specific regions of interest. WSI scans may be acquired at 2x, 10x, 20x, 40x, 60x and 100x, or a combination of magnifications. Thus the WSI pyramid allows the capabilities of interactive demonstrations with different magnifications, high-resolution (HR) display, and high performance classification based on imaging with special hematoxylin and eosin (H&E) tissue stains. The usage of color in digital pathology can demonstrate the pathological findings and give objective quantification measures that describe the characteristics of some spatial parts in tissues. The image pyramid consists of a base-level image of the highest magnification. Such a base-level image is obtained by successively aligning a series of smaller sub-images that are called tiles. The down-sampling is repeated successively until a thumbnail image is created at the lowest resolution level.

In recent years, digital images in medical histopathology have been explored in tasks of detection, segmentation, and classification of healthy and anomalous microscopic structures. Indeed, there is a large body of work in the literature about microscopic medical image analysis based on different types of features. Those features include texture, morphometric and intensity-based features, wavelets, co-occurrence and first-order statistics features, and features of color distribution over nuclei and cytoplasm, according to the H&E stain [7]. Basavanhallyet et al. employed one approach for WSI analysis where images are split and nuclear architecture and textural features are integrated from multiple fields of view (FOVs) in H&E stained breast cancer histopathology slides with various sizes. The developed system showed good differentiation between low, intermediate, and high-grade patients. Peikari et al. proposed a bag-of-words approach for cancer detection in WSI where image patches are randomly selected from whole tissue areas, patches are divided into smaller tiles, texture filters are applied to the tiles, histograms of filter responses are constructed, and statistical measures are derived from the histograms. The computed measures were used to train a support vector machine classifier to discriminate between clinically relevant and irrelevant patches. They report an area under the curve of 0. 91 for discriminating tumor and normal tissues. One of the major issues in WSI is the time and space complexities of pre-processing, segmentation, and classification. These complexities are due to the enormous size of a typical WSI image. One solution to alleviate this issue is the creation of a pyramidal structure where some tasks (e. g. object detection) can be performed on LR levels and the results are used to guide the completion of the task in higher-resolution levels. However, the performance outcomes on common tasks (e. g. anomaly detection or classification) for LR levels are typically worse than the outcomes on HR levels. Such complexity-performance trade-off can be reduced through transfer learning techniques. With the goal to transfer the knowledge from the source to target domain, recent developments in transfer learning have shown promising results for cross-domain recognition problems. Transfer learning can be applied in many applications from medical imaging diagnosis systems to material and texture classification.

For example, Zou and Yuen addressed the problem of classifying very low-resolution images (VLR) by a relationship learning framework where VLR-HR image patches and new data and discriminative constraints are exploited. Anam et al. boosted the classification performance of LR HEp-2 cell images using transfer learning among LR and HR images. In this paper, we employ a partial least-square (PLS) regression approach to learn the mapping from LR to HR WSI image patches. This mapping is used to improve the classification on LR WSI images. Moreover, we investigate the effect of noise on the mapping and subsequent classification. In particular, we simulated different scenarios where WSI images are contaminated with Gaussian noise. Three de-noising algorithms were applied, namely de-noising with the proposed PLS mapping, Block Matching 3D (BM3D) and the combination of PLS and BM3D. The experiments show the importance of the proposed transfer learning method in achieving better classification under LR noisy conditions.

Methodology

1. Dataset Description

The Camelyon16 dataset1 contains 400 WSIs of sentinel lymph node collected in Radboud University Medical Center (Nijmegen, the Netherlands), and the University Medical Center Utrecht (Utrecht, the Netherlands). We used 80 WSIs from Camelyon16 and divided them into 2983 patches. These correspond to 40 normal images (1476 patches) and 40 tumor images (1507patches). The patches were randomly divided into 1990 training patches and 994 testing ones. Open-Slide [14], a vendor-neutral C library, was used to read and manipulate digital slides of various formats. 1 https: //camelyon16. grand-challenge. org/ Various programming languages can be easily interfaced from the C library.

2. Feature extraction and dimensionality reduction

For each WSI image patch, we converted the color patch into a gray-scale one and then computed the multi-scale co-occurrence local binary pattern (MCLBP) [15]. MCLBP is a texture descriptor based on multi-scale co-occurrence LBP groups. All co-occurrence patterns (among LBP points) are rearranged into groups according to properties of the co-patterns, and encoding functions are extracted from each group. The major benefit of the MCLBP descriptor is that it can effectively capture strong correlation information between different LBP scales around the same central point and is also globally rotation invariant (GRI). A pattern for each point in each scale (LBPU) is calculated. After feature extraction, we normalized all types of features by feature scaling to avoid numerical issues. Then we used principal component analysis (PCA) to reduce the dimensionality of the MCLBP features.

3. Transfer learing across

WSI levels We utilize a transfer learning approach to infer LR-to-HR mapping between WSI image levels. Specifically, we investigated a method for learning the mapping of the LR image features where they are more correlated to the HR image features to enhance the classification performance by the partial least-square (PLS) regression algorithm [16]. The PLS is known as a regression model that projects the regressor and response matrices onto a low-dimensional latent linear subspace. The linear projections are chosen such that the covariance between latent scores of regressors (X) and responses (Y) is maximized. Then a linear mapping from the regressor latent scores to response latent scores is found. In this way, we apply PLS to learn linear projections that map the feature spaces into a common space in which they can be compared.

Results And Discussion

We ran several experiments to assess the performance of LR WSI image classification before and after the PLS-based mapping. The PLS mapping is used to map LR WSI features to HR WSI features and then apply an SVM classifier to the mapped features rather than the LR ones. Table III. 1 shows the results of the mappings between different LR and HR level pairs. Each result is the average of 10 repetitions and the data is shuffled each time. For each level pair, the table shows the classification accuracy on the HR level, LR level, and LR level with the PLS mapping. It is clear that the proposed PLS mapping algorithm improves the accuracy of LR WSI classification. It is noticed that the performance of the proposed mapping approach is affected by the proximity between the source LR level and the target HR level. Then, we tested the effect of noise on mapping and SVM classification.

1. noise-free high-resolution images,
2. noisy low-resolution images,
3. low-resolution images denoised with the PLS mapping method,
4. low-resolution images denoised with the BM3D method, and
5. low-resolution images denoised with a concatenation of the PLS mapping and BM3D methods.

Here are some remarks on the results. First of all, the accuracy for the HR images is constant as these images are free of noise. Secondly, the LR images show a clear drop in classification accuracy as the noise density increases. Thirdly, both of the PLS mapping and BM3D denoising methods lead to significant improvements in the robustness of the SVM classification against noise. Fourthly, applying the PLS mapping then BM3D denoising produce even better classification results.

Conclusions Ans Future Work

We have shown in this work a transfer-learning approach for boosting the cancer detection accuracy in whole slide images of sentinel lymph nodes. The proposed approaches offer the advantages of improved accuracy with low time and space costs. This is particularly important for WSI images that have increasingly very large sizes for higher-resolution levels. The proposed method also exhibited good performance under noise condition. In particular, the combination of the PLS-mapping and the BM3D denoising methods results in clear performance improvement over each individual method. For future work, we would like to investigate other learning methods and WSI descriptors to classify LR WSI images accurately and efficiently.