Angel Cruz-Roa | Universidad Nacional de Colombia (National University of Colombia) (original) (raw)

Journal Papers by Angel Cruz-Roa

The proliferative activity of breast tumors, which is routinely estimated by counting of mitotic ... more The proliferative activity of breast tumors, which is routinely estimated by counting of mitotic figures in hematoxylin and eosin stained histology sections, is considered to be one of the most important prognostic markers. However, mitosis counting is laborious, subjective and may suffer from low inter-observer agreement. With the wider acceptance of whole slide images in pathology labs, automatic image analysis has been proposed as a potential solution for these issues. In this paper, the results from the Assessment of Mitosis Detection Algorithms 2013 (AMIDA13) challenge are described. The challenge was based on a data set consisting of 12 training and 11 testing subjects, with more than one thousand annotated mitotic figures by multiple observers. Short descriptions and results from the evaluation of eleven methods are presented. The top performing method has an error rate that is comparable to the inter-observer agreement among pathologists.

Breast cancer (BCa) grading plays an important role in predicting disease aggressiveness and pati... more Breast cancer (BCa) grading plays an important role in predicting disease aggressiveness and patient outcome. A key component of BCa grade is mitotic count, which involves quantifying the number of cells in the process of dividing (i.e. undergoing mitosis) at a specific point in time. Currently mitosis counting is done manually by a pathologist looking at multiple high power fields on a glass slide under a microscope, an extremely laborious and time consuming process. The development of computerized systems for automated detection of mitotic nuclei, while highly desirable, is confounded by the highly variable shape and appearance of mitoses. Existing methods use either handcrafted features that capture certain morphological, statistical or textural attributes of mitoses or features learned with convolutional neural networks (CNN). While handcrafted features are inspired by the domain and the particular application, the data-driven CNN models tend to be domain agnostic and attempt to learn additional feature bases that cannot be represented through any of the handcrafted features. On the other hand, CNN is computationally more complex and needs a large number of labeled training instances. Since handcrafted features attempt to model domain pertinent attributes and CNN approaches are largely unsupervised feature generation methods, there is an appeal to attempting to combine these two distinct classes of feature generation strategies to create an integrated set of attributes that can potentially outperform either class of feature extraction strategies individually. In this paper, we present a cascaded approach for mitosis detection that intelligently combines a CNN model and handcrafted features (morphology, color and texture features). By employing a light CNN model, the proposed approach is far less demanding computationally, and the cascaded strategy of combining handcrafted features and CNN-derived features enables the possibility of maximizing performance by leveraging the disconnected feature sets. Evaluation on the public ICPR12 mitosis dataset that has 226 mitoses annotated on 35 High Power Fields (HPF, 400x magnification) by several pathologists and 15 testing HPFs yielded an F-measure of 0.7345. Our approach is accurate, fast and requires fewer computing resources compared to existent methods, making this feasible for clinical use.

Objective: The paper addresses the problem of finding visual patterns in histology image collecti... more Objective: The paper addresses the problem of finding visual patterns in histology image collections. In particular, it proposes a method for correlating basic visual patterns with high-level concepts combining an appropriate image collection representation with state-of-the-art machine learning techniques.
Methodology: The proposed method starts by representing the visual content of the collection using a bag-of-features strategy. Then, two main visual mining tasks are performed: finding associations between visual-patterns and high-level concepts, and performing automatic image annotation. Associations are found using minimum-redundancy-maximum-relevance feature selection and co-clustering analysis. Annotation is done by applying a support-vector-machine classifier. Additionally, the proposed method includes an interpretation mechanism that associates concept annotations with corresponding image regions.
The method was evaluated in two data sets: one comprising histology images from the different four fundamental tissues, and the other composed of histopathology images used for cancer diagnosis. Different visual-word representations and codebook sizes were tested. The performance in both concept association and image annotation tasks was qualitatively and quantitatively evaluated.
Results: The results show that the method is able to find highly discriminative visual features and to associate them to high-level concepts. In the annotation task the method showed a competitive performance: an increase of 21\% in f-measure with respect to the baseline in the histopathology data set, and an increase of 47\% in the histology data set.
Conclusions: The experimental evidence suggests that the bag-of-features representation is a good alternative to represent visual content in histology images. The proposed method exploits this representation to perform visual pattern mining from a wider perspective where the focus is the image collection as a whole, rather than individual images.

Histology images are an important resource for research, education and medical practice. The avai... more Histology images are an important resource for research, education and medical practice. The availability of image collections with reference purposes is limited to printed formats such as books and specialized journals. When histology image sets are published in digital formats, they are composed of some tens of images that do not represent the wide diversity of biological structures that can be found in fundamental tissues. Making a complete histology image collection available to the general public having a great impact on research and education in different areas such as medicine, biology and natural sciences. This work presents the acquisition process of a histology image collection with 20,000 samples in digital format, from tissue processing to digital image capturing. The main purpose of collecting these images is to make them available as reference material to the academic comunity. In addition, this paper presents the design and architecture of a system to query and explore the image collection, using content-based image retrieval tools and text-based search on the annotations provided by experts. The system also offers novel image visualization methods to allow easy identification of interesting images among hundreds of possible pictures. The system has been developed using a service-oriented architecture and allows web-based access in http://www.informed.unal.edu.co

Introducción: Los melanocitos epidérmicos están ampliamente separados entre sí, rodeados por un h... more Introducción: Los melanocitos epidérmicos están ampliamente separados entre sí, rodeados por un halo; son de citoplasma claro y núcleo picnótico, más pequeño que el de los queratocitos. En la cara es difícil diferenciar entre los cambios por exposición solar y un melanoma in situ, así como establecer si los bordes de resección de un melanoma in situ tienen tumor o si los melanocitos presentes sólo tienen cambios por el sol.
Objetivo: Cuantificar el número de melanocitos en adultos normales y en los bordes de resección sin tumor, de carcinomas basocelulares y de melanomas in situ de la piel malar.
Materiales y métodos: Se estudiaron veinticinco especímenes de piel tipo I-II de la mejilla de adultos mayores de 40 años, siete de autopsias de hombres, once de los bordes de carcinomas basocelulares y siete de los bordes de resección de melanomas in situ, libres de tumor. Con la coloración de hematoxilina-eosina, tres observadores contaron los melanocitos basales por milímetro lineal en cada espécimen, usando un fotomicroscopio Axiophot Zeiss.
Resultados: En un milímetro lineal (3 campos de 40X), el número de melanocitos fue de 18±3 en la piel normal, de 22±7 en los bordes del carcinoma basocelular y de 30±9 en los del melanoma in situ.
Conclusiones: El número máximo de melanocitos en un campo de 40X en los tejidos estudiados no debe exceder de 7,5±4 (30 melanocitos) por mm lineal. Un número mayor es una alerta que debe unirse a otros cambios para determinar si hay persistencia de melanoma in situ.

Conference Papers by Angel Cruz-Roa

This paper presents a software framework for large scale image processing and analysis over cloud... more This paper presents a software framework for large scale image processing and analysis over cloud computing resources. It adopts a decentralized and uncommitted resource allocation model where experimenters define their image processing pipelines in simple configuration files and add worker agents, on a casuistic and unscheduled manner, to contribute with their computing power to resolve the process load. Workers are Java processes encapsulated within virtual machines that only require access to a database where job definitions are stored. Each worker contains the neessary logic for it to act autonomously and yet contribute to a common task in an organized and effective manner, eliminating the need for a central scheduling node. This suits well research environments where access to computing resources is seldom or cannot be provisioned in advance, such as is often the case in Latin American research groups and institutions. We adopt a NoSQL storage model which fits nicely the simple data model that supports our framework while ensuring its storage scalability. For experimental evaluation we deployed sets of workers from our framework to perform image analysis tasks over the Amazon Cloud (AWS) using several of their elastic computing, storage and NoSQL services. As a result, our framework now enables researchers to run large scale image processing pipelines in an easy, affordable and unplanned manner with the capability to take over computing resources as they become available at run time. Moreover, experiments demonstrate its scalability and adaptability to efficiently exploit computing resources available on the Cloud at reasonable costs.

This paper presents BIGS the Big Image Data Analysis Toolkit, a software framework for large scal... more This paper presents BIGS the Big Image Data Analysis Toolkit, a software framework for large scale image processing and analysis over heterogeneous computing resources, such as those available in clouds, grids, computer clusters or throughout scattered computer resources (desktops, labs) in an opportunistic manner. Through BIGS, eScience for image processing and analysis is conceived to exploit coarse grained parallelism based on data partitioning and parameter sweeps, avoiding the need of inter-process communication and, therefore, enabling loosely coupled computing nodes (BIGS workers). It adopts an uncommitted resource allocation model where (1) experimenters define their image processing pipelines in a simple configuration file, (2) a schedule of jobs is generated and workers, as they become available, take over pending jobs as long as their dependency on other jobs is fulfilled. BIGS workers act autonomously, querying the job schedule to determine which one to take over. This removes the need for a central scheduling node, requiring only access by all workers to a shared information source. Furthermore, BIGS workers are encapsulated within different technologies to enable their agile deployment over the available computing resources. Currently they can be launched through the Amazon EC2 service over their cloud resources, through Java Web Start from any desktop computer and through regular scripting or SSH commands. This suits well different kinds of research environments, both when accessing dedicated computing clusters or clouds with committed computing capacity or when using opportunistic computing resources whose access is seldom or cannot be provisioned in advance. We also adopt a NoSQL storage model to ensure the scalability of the shared information sources required by all workers, packaging within BIGS implementations for HBase and Amazon's DynamoDB service. Overall, BIGS now enables researchers to run large scale image processing pipelines in an easy, affordable and unplanned manner with the capability to take over computing resources as they become available at run time. This is shown in this paper by using BIGS in different experimental setups in the Amazon cloud and in an opportunistic manner, demonstrating its configurability, adaptability and scalability capabilities.

Histopathological images are an important resource for clinical diagnosis and biomedical research... more Histopathological images are an important resource for clinical diagnosis and biomedical research. Automatic annotation of these images is particularly challenging from an image understanding point of view. This paper presents a new method for automatic histopathological image annotation based on three complementary strategies, first, a part-based image representation, called the bag of features, which takes advantage of the natural redundancy of histopathological images, second, a latent topic model, based on non-negative matrix factorization, which is in charge of capturing the high-level visual patterns, and, third, a probabilistic annotation model that connects visual patterns with the histopathological image semantics. The method was evaluated using 1, 604 annotated images of basal cell carcinoma, a collection with different types of skin cancer. The preliminary results demonstrate an improvement on precision and recall of 24 % and 64 % against a baseline annotation method based on support vector machines.

This paper presents a framework to analyse visual patterns in a collection of medical images in a... more This paper presents a framework to analyse visual patterns in a collection of medical images in a two stage procedure. First, a set of representative visual patterns from the image collection is obtained by constructing a visual-word dictionary under a bag-of-features approach. Second, an analysis of the relationships between visual patterns and semantic concepts in the image collection is performed. The most important visual patterns for each semantic concept are identified using correlation analysis. A matrix visualization of the structure and organization of the image collection is generated using a cluster analysis. The experimental evaluation was conducted on a histopathology image collection and results showed clear relationships between visual patterns and semantic concepts, that in addition, are of easy interpretation and understanding.

A method for automatic analysis and interpretation of histopathology images is presented. The met... more A method for automatic analysis and interpretation of histopathology images is presented. The method uses a representation of the image data set based on bag of features histograms built from visual dictionary of Haar-based patches and a novel visual latent semantic strategy for characterizing the visual content of a set of images. One important contribution of the method is the provision of an interpretability layer, which is able to explain a particular classification by visually mapping the most important visual patterns associated with such classification. The method was evaluated on a challenging problem involving automated discrimination of medulloblastoma tumors based on image derived attributes from whole slide images as anaplastic or non-anaplastic. The data set comprised 10 labeled histopathological patient studies, 5 for anaplastic and 5 for non-anaplastic, where 750 square images cropped randomly from cancerous region from whole slide per study. The experimental results show that the new method is competitive in terms of classification accuracy achieving 0.87 in average.

This paper proposes an adaptive image representation learning method for cervix cancer tumor dete... more This paper proposes an adaptive image representation learning method for cervix cancer tumor detection. The method learns the representation in two stages, a local feature description using a sparse dictionary learning and a global image representation using a bag-of-features (BOF) approach. The resultant representation is thus a BOF histogram, learned from a sparse local patch representation. The parameters of the sparse learning representation algorithm are tuned up by searching dictionaries with low coherence and high sparsity. The proposed method was evaluated in a dataset of 394 cervical histology images with tumoral and non-tumoral pathologies acquired at a 10X magnification and a resolution of 3800 × 3000 pixels in RGB color. A conventional BOF image representation, using a linearized raw-block patch descriptor, was selected as the baseline. The preliminary results show that our proposed method improves the baseline for all different BOF dictionary sizes (125, 250, 500, 1000 and 2000). Under a 10 cross-validation test and a 2000 BOF dictionary, the best performance was 0.77 ± 0.04 in average accuracy, improving in 2.5% the baseline. These results suggest that a learning-from-data approach could be used in different stages of an image classifier construction pipeline, in particular for the image representation stage.

Image representation is an important issue for medical image analysis, classification and retriev... more Image representation is an important issue for medical image analysis, classification and retrieval. Recently, the bag of features approach has been proposed to classify natural scenes, using an analogy in which visual features are to images as words are to text documents. This process involves feature detection and description, construction of a visual vocabulary and image representation building through visualword occurrence analysis. This paper presents an evaluation of different representations obtained from the bag of features approach to classify histopathology images. The obtained image descriptors are processed using appropriate kernel functions for Support Vector Machines classifiers. This evaluation includes extensive experimentation of different strategies, and analyses the impact of each configuration in the classification result.

This paper describes an extension of the Bag-of-Visual-Words (BoVW) representation for image clas... more This paper describes an extension of the Bag-of-Visual-Words (BoVW) representation for image classification (IC) of histophatology images. This representation is one of the most used approaches in several high-level computer vision tasks. However, the BoVW representation has an important limitation: the disregarding of spatial information among visual words. This information may be useful to capture discriminative visual-patterns in specific computer vision tasks. In order to overcome this problem we propose the use of visual n-grams. N-grams based-representations are very popular in the field of natural language processing (NLP), in particular within text mining and information retrieval. We propose building a codebook of n-grams and then representing images by histograms of visual n-grams. We evaluate our proposal in the challenging task of classifying histopathology images. The novelty of our proposal lies in the fact that we use n-grams as attributes for a classification model (together with visual-words, i.e., 1-grams). This is common practice within NLP, although, to the best of our knowledge, this idea has not been explored yet within computer vision. We report experimental results in a database of histopathology images where our proposed method outperforms the traditional BoVWs formulation.

This paper presents a novel method for basal-cell carcinoma detection, which combines state-of-th... more This paper presents a novel method for basal-cell carcinoma detection, which combines state-of-the-art methods for unsupervised feature learning (UFL) and bag of features (BOF) representation. BOF, which is a form of representation learning, has shown a good performance in automatic histopathology image classification. In BOF, patches are usually represented using descriptors such as Scale-Invariant Feature Transform (SIFT) and Discrete Cosine Transformation (DCT). We propose to use UFL to learn the patch representation itself. This is accomplished by applying a topographic UFL method (T-RICA), which automatically learns visual invariance properties of color, scale and rotation from an image collection. These learned features also reveals these visual properties associated to cancerous and healthy tissues and improves carcinoma detection results by 7% with respect to traditional autoencoders, and 6% with respect to standard DCT representations obtaining in average 92% in terms of F-score and 93% of balanced accuracy.

Breast cancer (BCa) grading plays an important role in predicting disease aggressiveness and pati... more Breast cancer (BCa) grading plays an important role in predicting disease aggressiveness and patient outcome. A key component of BCa grade is mitotic count, which involves quantifying the number of cells in the process of dividing (i.e. undergoing mitosis) at a specific point in time. Currently mitosis counting is done manually by a pathologist looking at multiple high power fields on a glass slide under a microscope, an extremely laborious and time consuming process. The development of computerized systems for automated detection of mitotic nuclei, while highly desirable, is confounded by the highly variable shape and appearance of mitoses. Existing methods use either handcrafted features that capture certain morphological, statistical or textural attributes of mitoses or features learned with convolutional neural networks (CNN). While handcrafted features are inspired by the domain and the particular application, the data-driven CNN models tend to be domain agnostic and attempt to learn additional feature bases that cannot be represented through any of the handcrafted features. On the other hand, CNN is computationally more complex and needs a large number of labeled training instances. Since handcrafted features attempt to model domain pertinent attributes and CNN approaches are largely unsupervised feature generation methods, there is an appeal to attempting to combine these two distinct classes of feature generation strategies to create an integrated set of attributes that can potentially outperform either class of feature extraction strategies individually. In this paper, we present a cascaded approach for mitosis detection that intelligently combines a CNN model and handcrafted features (morphology, color and texture features). By employing a light CNN model, the proposed approach is far less demanding computationally, and the cascaded strategy of combining handcrafted features and CNN-derived features enables the possibility of maximizing performance by leveraging the disconnected feature sets. Evaluation on the public ICPR12 mitosis dataset that has 226 mitoses annotated on 35 High Power Fields (HPF, x400 magnification) by several pathologists and 15 testing HPFs yielded an F-measure of 0.7345. Apart from this being the second best performance ever recorded for this MITOS dataset, our approach is faster and requires fewer computing resources compared to extant methods, making this feasible for clinical use.

This paper presents a deep learning approach for automatic detection and visual analysis of invas... more This paper presents a deep learning approach for automatic detection and visual analysis of invasive ductal carcinoma (IDC) tissue regions in whole slide images (WSI) of breast cancer (BCa). Deep learning approaches are learn-from-data methods involving computational modeling of the learning process. This approach is similar to how human brain works using different interpretation levels or layers of most representative and useful features resulting into a hierarchical learned representation. These methods have been shown to outpace traditional approaches of most challenging problems in several areas such as speech recognition and object detection. Invasive breast cancer detection is a time consuming and challenging task primarily because it involves a pathologist scanning large swathes of benign regions to ultimately identify the areas of malignancy. Precise delineation of IDC in WSI is crucial to the subsequent estimation of grading tumor aggressiveness and predicting patient outcome. DL approaches are particularly adept at handling these types of problems, especially if a large number of samples are available for training, which would also ensure the generalizability of the learned features and classifier. The DL framework in this paper extends a number of convolutional neural networks (CNN) for visual semantic analysis of tumor regions for diagnosis support. The CNN is trained over a large amount of image patches (tissue regions) from WSI to learn a hierarchical part-based representation. The method was evaluated over a WSI dataset from 162 patients diagnosed with IDC. 113 slides were selected for training and 49 slides were held out for independent testing. Ground truth for quantitative evaluation was provided via expert delineation of the region of cancer by an expert pathologist on the digitized slides. The experimental evaluation was designed to measure classifier accuracy in detecting IDC tissue regions in WSI. Our method yielded the best quantitative results for automatic detection of IDC regions in WSI in terms of F-measure and balanced accuracy (71.80%, 84.23%), in comparison with an approach using handcrafted image features (color, texture and edges, nuclear textural and architecture), and a machine learning classifier for invasive tumor classification using a Random Forest. The best performing handcrafted features were fuzzy color histogram (67.53%, 78.74%) and RGB histogram (66.64%, 77.24%). Our results also suggest that at least some of the tissue classification mistakes (false positives and false negatives) were less due to any fundamental problems associated with the approach, than the inherent limitations in obtaining a very highly granular annotation of the diseased area of interest by an expert pathologist.

This paper presents and evaluates a deep learning architecture for automated basal cell carcinoma... more This paper presents and evaluates a deep learning architecture for automated basal cell carcinoma cancer detection that integrates (1) image representation learning, (2) image classification and (3) result interpretability. A novel characteristic of this approach is that it extends the deep learning architecture to also include an interpretable layer that highlights the visual patterns that contribute to discriminate between cancerous and normal tissues patterns, working akin to a digital staining which spotlights image regions important for diagnostic decisions. Experimental evaluation was performed on set of 1,417 images from 308 regions of interest of skin histopathology slides, where the presence of absence of basal cell carcinoma needs to be determined. Different image representation strategies, including bag of features (BOF), canonical (discrete cosine transform (DCT) and Haar-based wavelet transform (Haar)) and proposed learned-from-data representations, were evaluated for comparison. Experimental results show that the representation learned from a large histology image data set has the best overall performance (89.4% in F-measure and 91.4% in balanced accuracy), which represents an improvement of around 7% over canonical representations and 3% over the best equivalent BOF representation.

Breast cancer (BCa) grading plays an important role in predicting disease aggressiveness and pati... more Breast cancer (BCa) grading plays an important role in predicting disease aggressiveness and patient outcome. A key component of BCa grade is mitotic count, which involves quantifying the number of cells in the process of dividing (i.e. undergoing mitosis) at a specific point in time. Currently mitosis counting is done manually by a pathologist looking at multiple high power fields on a glass slide under a microscope, an extremely laborious and time consuming process. The development of computerized systems for automated detection of mitotic nuclei, while highly desirable, is confounded by the highly variable shape and appearance of mitoses. Existing methods use either handcrafted features that capture certain morphological, statistical or textural attributes of mitoses or features learned with convolutional neural networks (CNN). While handcrafted features are inspired by the domain and the particular application, the data-driven CNN models tend to be domain agnostic and attempt to learn additional feature bases that cannot be represented through any of the handcrafted features. On the other hand, CNN is computationally more complex and needs a large number of labeled training instances. Since handcrafted features attempt to model domain pertinent attributes and CNN approaches are largely unsupervised feature generation methods, there is an appeal to attempting to combine these two distinct classes of feature generation strategies to create an integrated set of attributes that can potentially outperform either class of feature extraction strategies individually. In this paper, we present a cascaded approach for mitosis detection that intelligently combines a CNN model and handcrafted features (morphology, color and texture features). By employing a light CNN model, the proposed approach is far less demanding computationally, and the cascaded strategy of combining handcrafted features and CNN-derived features enables the possibility of maximizing performance by leveraging the disconnected feature sets. Evaluation on the public ICPR12 mitosis dataset that has 226 mitoses annotated on 35 High Power Fields (HPF, x400 magnification) by several pathologists and 15 testing HPFs yielded an F-measure of 0.7345. Apart from this being the second best performance ever recorded for this MITOS dataset, our approach is faster and requires fewer computing resources compared to extant methods, making this feasible for clinical use.