Skip to main content
Fig. 3 | BMC Bioinformatics

Fig. 3

From: Marker-controlled watershed with deep edge emphasis and optimized H-minima transform for automatic segmentation of densely cultivated 3D cell nuclei

Fig. 3

Segmentation of nuclei via system configurations with a demonstration using the configuration \(\{C, M_{2DE}, S\}\). All the configurations utilise an input volume that is expanded near isotropic (1). The expanded volume is transformed with a chosen U-Net model type (\(U_{M_{3D}}, U_{M_{3DE}}, U_{M_{2DE}}\) or \(U_{S}\)) into one of the nuclei masks \(M \in \{M_{3D}, M_{3DE}, M_{2DE}\}\), where \(M_{3D}\) denotes 3D nuclei mask, \(M_{3DE}\) 3D edge emphasizing nuclei mask and \(M_{2DE}\) 2D edge emphasizing nuclei mask, and optionally to binary seeds S (2). Instance segmentation is performed using one of the three different marker-controlled watershed methods A, B or C (3). The method A transforms binary seeds into markers via connected component (CC) analysis, and feeds markers and nuclei mask to the marker-controlled watershed transform, \(WS_m\), which computes distance transform (DT) of nuclei mask and creates an instance segmentation. The method B uses H-minima-based marker-controlled watershed, \(WS_h\), which input consist of nuclei mask and a h-value. Markers are determined from the nuclei mask via DT and H-minima transform, and similarly as in \(WS_m\), DT and markers are transformed into an instance segmentation. The method C is otherwise the same as A but generates markers by feeding seeds to \(WS_h\). Given a mask, optionally seeds and N different h-values, a chosen watershed method produces N different segmentation maps \(O_i\). The segmentation \(O_i\) with the highest average roundness score \(\phi _R\) is chosen as the final segmentation (3)

Back to article page