PhD student, Univerisity of Alberta
Tumor cell detection is of great interest to a wide range of medical scenarios. An example is the diagnosis and treatment of breast cancer, which is one of the most common female diseases leading to death worldwide. The number of proliferating (e.g. Ki67 positive) tumor cells is an important index associated with the severity of disease clinically.
One available method involves detection of the nuclei of proliferating cells using traditional image analysis techniques on a microscopic image. However, it has been proven to be challenging because of inability to distinguish tumor cells from surrounding normal tissue like vessels, fat and fibrous tissue, especially in reality the resolution of input medical image could be very high, at the same time the target cells could easily be extremely dense.
The goal of this research is to create a computer vision tool that could actually detect and count the number of tumor nuclei and then give a percentage of Ki67 positive tumor nuclei over the total number of tumor nuclei.
In the technical side, we propose to integrate compressed sensing (CS) with convolutional neural network (CNN) for cell detection and localization. There are two principal reasons behind this merging. The first one is rather pragmatic that one would strive to find an end-to-end training system that would make the training procedure accurate and straightforward. The second reason is technical. CS will let us turn the detection of variable number of objects into a fixed length vector regression task, where one can apply several state-of-the-art CNN architectures. The proposed method is evaluated with several state-of-the-art approaches on three public-available cell datasets and obtains more superior performances.
Abstract: 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.
Pub.: 31 Dec '14, Pinned: 30 Jun '17
Abstract: The lack of publicly available ground-truth data has been identified as the major challenge for transferring recent developments in deep learning to the biomedical imaging domain. Though crowdsourcing has enabled annotation of large scale databases for real world images, its application for biomedical purposes requires a deeper understanding and hence, more precise definition of the actual annotation task. The fact that expert tasks are being outsourced to non-expert users may lead to noisy annotations introducing disagreement between users. Despite being a valuable resource for learning annotation models from crowdsourcing, conventional machine-learning methods may have difficulties dealing with noisy annotations during training. In this manuscript, we present a new concept for learning from crowds that handle data aggregation directly as part of the learning process of the convolutional neural network (CNN) via additional crowdsourcing layer (AggNet). Besides, we present an experimental study on learning from crowds designed to answer the following questions: (i) Can deep CNN be trained with data collected from crowdsourcing?, (ii) How to adapt the CNN to train on multiple types of annotation datasets (ground truth and crowd-based)?, (iii) How does the choice of annotation and aggregation affect the accuracy? Our experimental setup involved Annot8, a self-implemented web-platform based on Crowdflower API realizing image annotation tasks for a publicly available biomedical image database. Our results give valuable insights into the functionality of deep CNN learning from crowd annotations and prove the necessity of data aggregation integration.
Pub.: 19 Feb '16, Pinned: 30 Jun '17
Abstract: One example apparatus associated with detecting mitosis in breast cancer pathology images by combining handcrafted (HC) and convolutional neural network (CNN) features in a cascaded architecture includes a set of logics that acquires an image of a region of tissue, partitions the image into candidate patches, generates a first probability that the patch is mitotic using an HC feature set and a second probability that the patch is mitotic using a CNN-learned feature set, and classifies the patch based on the first probability and the second probability. If the first and second probabilities do not agree, the apparatus trains a cascaded classifier on the CNN-learned feature set and the HC feature set, generates a third probability that the patch is mitotic, and classifies the patch based on a weighted average of the first probability, the second probability, and the third probability.
Pub.: 30 Aug '16, Pinned: 30 Jun '17
Abstract: To diagnose breast cancer, the number of mitotic cells present in histology sections is an important indicator for examining and grading biopsy specimen. This study aims at improving the accuracy of automated mitosis detection by characterizing mitotic cells in wavelet based multi-resolution representations via a non-Gaussian modeling method. The potential mitosis candidates were decomposed into multi-scale forms by an undecimated dual-tree complex wavelet transform. Two non-Gaussian models (the generalized Gaussian distribution (GGD) and the symmetric alpha-stable (SαS) distributions) were used to accurately model the heavy-tailed behavior of wavelet marginal distributions. The method was evaluated on two independent data cohorts, including the benchmark dataset (MITOS), via a support vector machine classifier. The quantitative results shows that the bivariate SαS model achieved superior classification performance with the area under the curve value of 0.82 in comparison with 0.79 for bivariate GGD, 0.77 for univariate SαS, 0.72 for univariate GGD, and 0.59 for Gaussian model. Since both mitotic and non-mitotic cells appear as small objects with a large variety of shapes, characterization of mitosis is a hard problem. The inter-scale dependencies of wavelet coefficients allowing extraction of salient features within the cells that are more likely to appear at all different scales were captured by the bivariate non-Gaussian models, leading to more accurate detection results. The presented automated mitosis detection method might assist pathologists in enhancing the operational efficiency and productivity as well as improving diagnostic confidence.
Pub.: 06 Jan '17, Pinned: 30 Jun '17