Cell detection in bright-field microscope images is a challenging task due to the low contrast delivered by the simple bright-field hardware. In addition, cells usually show a great diversity in size, orientation, and visual appearance. In order to cope with these difficulties, we developed a learning-based system for cell detection. It was designed with the following goals in mind: First, invariance to cell size, cell orientation, and illumination conditions. Second, the system should make no prior assumptions about cell shape. Third, no manual tuning of the system parameters is required. Fourth, the result of detection should be well-suited for tracking. The latter is particularly important for time-lapse microscopy.

The algorithm pipeline is shown in the figure: (A) A defocused image is used as input. (B) Scale invariant feature transform (SIFT) is applied. (C) Features are extracted for each SIFT keypoint and fed into a random forest classifier in order to classify the keypoints into cell and background. (D) Background keypoints are discarded and intensity profiles are extracted between each two “nearby” keypoints. (E) Features are extracted for each profile and fed into another random forest classifier in order to classify the profiles into “inner” and “cross” profiles. Any profile between two keypoints in the same cell is called inner profile, otherwise it is called cross profile. (F) The result of the previous classification step is probabilistic. The probability that a profile between two keypoints is an inner profile is considered as a similarity measure between these two keypoints. The matrix of similarity measure values for all cell keypoints is used as input for agglomerative hierarchical clustering. A customized linkage method was developed and used for the clustering in order to incorporate application-specific information. (G) A weighted average of the keypoint coordinates inside each cluster is computed and considered as a cell center.  

The keypoint features and the profile features were carefully designed in order to make them, as far as possible, invariant to scale, orientation, and illumination conditions.

The evaluation database contained more than 3500 manually labeled cells in three real cell lines and 30000 cells in two simulated cell lines. The detection error was in the best case close to zero, and in the worst case less than 15%.

The algorithm was developed for and evaluated on bright-field microscope images. However, due to its general design, we also used it for molecule detection in scanning tunneling microscope images.

Source code: coming soon.

In this work we propose a set of features for stenosis classification. The proposed processing pipeline starts with segmenting a digital subtraction angiography (DSA) volume using a level set segmentation framework. The resulting mask M is skeletonized using a distance-transform based skeletonization method. The skeleton tree is then traced and segmented into branches. Bifurcation points are detected and excluded from further processing because the morphology of vascular trees inside bifurcations is inherently different from the morphology inside branches. At each medial axis point, whose distance to the closest bifurcation is above a certain threshold, a cross-section of the binary mask M is extracted perpendicular to the medial axis (figure (A)). Then a feature vector is extracted and fed into a support vector machine (SVM) in order to classify the point as belonging to a healthy or a stenotic branch.

Figure (B) shows one of the stenosis features which proved to be discriminative between stenotic and healthy medial axis points. It is the second derivative of the cross-section's circumference as a function of the abscissa along the medial axis.

The evaluation was performed on 12 healthy and 4 stenotic branches extracted from 6 DSA volumes. In addition, 7 simulated stenotic branches were used.  In total, the dataset contained 23 branches. The simulated branches were investigated by a medical doctor and given an average grade of 65 % in terms of similarity to natural stenoses. The total number of feature vectors extracted from all branches was 189 stenotic feature vectors and 3087 healthy feature vector. Figure (C) shows a Sammon map projection of this feature space. In order to cope with data imbalance, a control-sensitivity SVM model was used. In a leave-one-branch-out cross validation, the misclassification rate was 6.7 % of the feature vectors using a radial basis function kernel and 6.4 % using a linear kernel.