Typical applications of zebrafish segmentation. a Fluorescence images visualisation and evaluation. Bright-field zebrafish images offer reference for the shape of the specimen (column one). Fluorescent images present informative signals, e.g. the blood vessels in green (column two). Accurate segmentation of the bright-field image provides a good shape reference to evaluate the fluorescent signals, for example, the development and concentration of specific cells (column three). b 3D zebrafish reconstruction from axial views. Axial-view zebrafish images (column one) are segmented to obtain 2D binary shapes (column two), from which the axial-view-based 3D reconstruction produces 3D models as well as 3D measurements (column three) (colour figure online)

Segmentations by different methods for a zebrafish specimen in lateral position. Blue bounding box indicates the expected segmentations, and red bounding box indicates inaccurate segmentations. a Segmentation by the geodesic active contours (GAC) model. Due to the edge sensitivity, the GAC model fails to detect the tail of the specimen. b Segmentation by Chan–Vese (CV) model. The partial transparency of the specimen makes it difficult for a region-based method to discriminate the object from the background. c Segmentation by a local region-based level set (LRLS) model. Similar problem occurs that the tail of the specimen is incorrectly segmented. d Segmentation by an improved level set (ILS) method. e Segmentation by mean shift (MS) algorithm. Better results are obtained though; edge sensitivity becomes worse. f Segmentation by the proposed hybrid (HY) method. The accurate segmentation presents a natural and compact shape description for the zebrafish specimen (colour figure online)

A pipeline schematic of the hybrid method. a MS algorithm is applied to improve the visibility of the transparent regions and weak edges. b An enclosed contour is extracted from the segmentation candidate in (a). c A distance-regularised level set function (LSF) is initialised from the zebrafish contour in (b). d The ILS method is activated and applied on the original image. e Another segmentation candidate is generated. f An initial hybrid segmentation of the zebrafish is obtained by stitching the remarkable segments from the two candidates according to pre-defined protocols. g A refinement is followed to fine-tune the segmentation which can accurately represent the shape of the zebrafish

Segmentation results visualisation of different methods on one zebrafish example from Dataset A. The object is positioned in ventral. GAC = geodesic active contours model [9]. LRLS = local region-based level set model [11]. ILS = improved level set method [12]. MS = mean shift algorithm [13]. HY = the proposed hybrid method. FCN = fully convolutional neural networks. GT = groundtruth

Segmentation results visualisation of different methods on one zebrafish example from Dataset A. The object is in titled position

Segmentation results visualisation of different methods on one zebrafish example from Dataset A. The object is positioned in lateral

Segmentation results visualisation of different methods on one zebrafish example from Dataset B. The object is positioned in ventral

Segmentation results visualisation of different methods on one zebrafish example from Dataset B. The object is in tilted position

Segmentation results visualisation of different methods on one zebrafish example from Dataset B. The object is positioned in lateral

Segmentation results visualisation on Dataset A. We randomly select one subject from each of the three larval stages and show the lateral view. In each bounding box, the upper figure shows the result obtained from the FCN trained from our annotated subset of Dataset A. And the bottom figure shows the result obtained from our HY method