Improved breast cancer histological grading using deep learning.

BACKGROUND: The Nottingham histological grade (NHG) is a well-established prognostic factor for breast cancer that is broadly used in clinical decision making. However, ∼50% of patients are classified as grade 2, an intermediate risk group with low clinical value. To improve risk stratification of NHG 2 breast cancer patients, we developed and validated a novel histological grade model (DeepGrade) based on digital whole-slide histopathology images (WSIs) and deep learning. PATIENTS AND METHODS: In this observational retrospective study, routine WSIs stained with haematoxylin and eosin from 1567 patients were utilised for model optimisation and validation. Model generalisability was further evaluated in an external test set with 1262 patients. NHG 2 cases were stratified into two groups, DG2-high and DG2-low, and the prognostic value was assessed. The main outcome was recurrence-free survival. RESULTS: DeepGrade provides independent prognostic information for stratification of NHG 2 cases in the internal test set, where DG2-high showed an increased risk for recurrence (hazard ratio [HR] 2.94, 95% confidence interval [CI] 1.24-6.97, P = 0.015) compared with the DG2-low group after adjusting for established risk factors (independent test data). DG2-low also shared phenotypic similarities with NHG 1, and DG2-high with NHG 3, suggesting that the model identifies morphological patterns in NHG 2 that are associated with more aggressive tumours. The prognostic value of DeepGrade was further assessed in the external test set, confirming an increased risk for recurrence in DG2-high (HR 1.91, 95% CI 1.11-3.29, P = 0.019). CONCLUSIONS: The proposed model-based stratification of patients with NHG 2 tumours is prognostic and adds clinically relevant information over routine histological grading. The methodology offers a cost-effective alternative to molecular profiling to extract information relevant for clinical decisions.

Morphology-guided graph search for untangling objects: C. elegans analysis.

We present a novel approach for extracting cluttered objects based on their morphological properties. Specifically, we address the problem of untangling Caenorhabditis elegans clusters in high-throughput screening experiments. We represent the skeleton of each worm cluster by a sparse directed graph whose vertices and edges correspond to worm segments and their adjacencies, respectively. We then search for paths in the graph that are most likely to represent worms while minimizing overlap. The worm likelihood measure is defined on a low-dimensional feature space that captures different worm poses, obtained from a training set of isolated worms. We test the algorithm on 236 microscopy images, each containing 15 C. elegans worms, and demonstrate successful cluster untangling and high worm detection accuracy.

Quantification of colocalization and cross-talk based on spectral angles.

Common methods for quantification of colocalization in fluorescence microscopy typically require cross-talk free images or images where cross-talk has been eliminated by image processing, as they are based on intensity thresholding. Quantification of colocalization includes not only calculating a global measure of the degree of colocalization within an image, but also a classification of each image pixel as showing colocalized signals or not. In this paper, we present a novel, automated method for quantification of colocalization and classification of image pixels. The method, referred to as SpecDec, is based on an algorithm for spectral decomposition of multispectral data borrowed from the field of remote sensing. Pixels are classified based on hue rather than intensity. The hue distribution is presented as a histogram created by a series of steps that compensate for the quantization noise always present in digital image data, and classification rules are thereafter based on the shape of the angle histogram. Detection of colocalized signals is thus only dependent on the hue, making it possible to classify also low-intensity objects, and decoupling image segmentation from detection of colocalization. Cross-talk will show up as shifts of the peaks of the histogram, and thus a shift of the classification rules, making the method essentially insensitive to cross-talk. The method can also be used to quantify and compensate for cross-talk, independent of the microscope hardware.

BlobFinder, a tool for fluorescence microscopy image cytometry.

Images can be acquired at high rates with modern fluorescence microscopy hardware, giving rise to a demand for high-speed analysis of image data. Digital image cytometry, i.e., automated measurements and extraction of quantitative data from images of cells, provides valuable information for many types of biomedical analysis. There exists a number of different image analysis software packages that can be programmed to perform a wide array of useful measurements. However, the multi-application capability often compromises the simplicity of the tool. Also, the gain in speed of analysis is often compromised by time spent learning complicated software. We provide a free software called BlobFinder that is intended for a limited type of application, making it easy to use, easy to learn and optimized for its particular task. BlobFinder can perform batch processing of image data and quantify as well as localize cells and point like source signals in fluorescence microscopy images, e.g., from FISH, in situ PLA and padlock probing, in a fast and easy way.

Combining intensity, edge and shape information for 2D and 3D segmentation of cell nuclei in tissue sections.

We present a region-based segmentation method in which seeds representing both object and background pixels are created by combining morphological filtering of both the original image and the gradient magnitude of the image. The seeds are then used as starting points for watershed segmentation of the gradient magnitude image. The fully automatic seeding is done in a generous fashion, so that at least one seed will be set in each foreground object. If more than one seed is placed in a single object, the watershed segmentation will lead to an initial over-segmentation, i.e. a boundary is created where there is no strong edge. Thus, the result of the initial segmentation is further refined by merging based on the gradient magnitude along the boundary separating neighbouring objects. This step also makes it easy to remove objects with poor contrast. As a final step, clusters of nuclei are separated, based on the shape of the cluster. The number of input parameters to the full segmentation procedure is only five. These parameters can be set manually using a test image and thereafter be used on a large number of images created under similar imaging conditions. This automated system was verified by comparison with manual counts from the same image fields. About 90% correct segmentation was achieved for two- as well as three-dimensional images.