Method for accurate segmentation of unstained living adherent cells in differential interference contrast images
In cell instance segmentation of an original cell image obtained by microscopy, the original cell image is rotated by a rotation angle to yield a rotated cell image. A first machine-learning model for cell instance segmentation processes the original and rotated cell images such that effectively, horizontal and oblique boxes are used in bounding cells of the original cell image for enhancing cell-segmentation accuracy. A set of predictions is generated for each cell instance identified from the original and rotated cell images, yielding a plurality of sets of predictions for all cell instances. The plurality of sets of predictions is processed to remove any set having predicted space not simply connected, any unwanted set classified by a second machine-learning model as a poor indicator according to a certain performance criterion, and any redundant set. The plurality of sets of predictions then produces a segmentation map for the original cell image.
1. A computer-implemented method of cell instance segmentation of an original cell image obtained by microscopy, the method comprising:
rotating the original cell image by a rotation angle to yield a rotated cell image;
using a first machine-learning model configured for cell instance segmentation to process the original and rotated cell images such that effectively, horizontal boxes and oblique boxes are used in bounding cells of the original cell image for enhancing cell-segmentation accuracy in segmenting the original cell image, wherein the first machine-learning model generates a set of predictions for an individual cell instance in a plurality of cell instances identified from the original and rotated cell images such that a first plurality of sets of predictions is obtained for the plurality of cell instances, and wherein the set of predictions includes a bounding box for bounding the individual cell instance, and a foreground-background binary mask for predicting a space occupied by the individual cell instance;
removing, from the first plurality of sets of predictions, any set of predictions having the predicted space not simply connected to thereby form a second plurality of sets of predictions;
using a second machine-learning model to classify an individual set of predictions in the second plurality of sets of predictions as a wanted set of predictions or an unwanted set of predictions according to a prediction performance of the individual set of predictions;
removing, from the second plurality of sets of predictions, any unwanted set of predictions to form a third plurality of sets of predictions;
removing, from the third plurality of sets of predictions, any redundant set of predictions to form a fourth plurality of sets of predictions, wherein an individual redundant set of predictions as generated for a cell instance located or equivalently located at a location on the original cell image has another set of predictions generated for the same location in the third plurality of sets of predictions; and
using the fourth plurality of sets of predictions to produce a segmentation map for the original cell image.
2. The method of claim 1 , wherein the rotation angle is 45°.
3. The method of claim 1 , wherein the set of predictions further includes one or more class labels for classifying the individual cell instance.
4. The method of claim 3 , wherein the first machine-learning model for processing an input image comprises:
a backbone for extracting features of the input image to generate feature maps;
a region proposal network (RPN) for generating region proposals from the feature maps;
a RoIAlign module for aligning the input image and the feature maps;
a plurality of fully-connected (FC) layers for bounding-box generation and class-label generation from the feature maps; and
a plurality of convolution layers for foreground-background binary mask generated from the feature maps.
5. The method of claim 4 , wherein the backbone is implemented as ResNet-50 or ResNet-101.
6. The method of claim 1 , wherein the second machine-learning model is a ResNet classifier.
7. The method of claim 1 , wherein the cells of the original cell image are adherent cells.
8. The method of claim 1 , wherein the cells of the original cell image are unstained.
9. The method of claim 1 , wherein the cells of the original cell image are living cells.
10. The method of claim 1 , wherein the original cell image is a differential interference contrast (DIC) image.
11. A non-transitory computer readable storage medium stored with program instructions which, when executed by a computing processor, cause the computing processor to execute a process of segmenting cells in original cell images according to the method of claim 1 .
12. A computer-implemented method for segmenting cells in plural original cell images obtained by microscopy, the method comprising the steps of:
(a) rotating the original cell images to obtain rotated cell images;
(b) training a deep-learning-based instance segmentation algorithm on a cell image dataset comprising the original cell images and rotated cell images;
(c) using the trained deep-learning-based instance segmentation algorithm to process the original cell images and rotated cell images to obtain predictions, wherein the predictions include predicted bounding boxes, classification labels, and foreground-background binary masks;
(d) concatenating the predictions for every pair of an individual original cell image and a corresponding rotated cell image throughout the entire dataset;
(e) filtering the concatenated predictions as identified in the step (d);
(f) saving the filtered predictions offline;
(g) compacting the saved offline predictions by removing respective saved offline predictions having corresponding foreground-background binary masks not simply connected;
(h) formulating masked images for the saved offline predictions identified in the step (g), wherein an individual masked image is obtained by overlaying a background part of corresponding predicted binary mask on a corresponding original image;
(i) training a deep-learning-based binary classification algorithm with the masked images obtained in the step (h) to generate unwanted and wanted masks; and
(j) processing the wanted masks obtained in the step (i) to produce final segmentation.
13. The method of claim 12 , wherein the cells are adherent cells.
14. The method of claim 12 , wherein the cells are unstained.
15. The method of claim 12 , wherein the cells are living cells.
16. The method of claim 12 , wherein the original cell images are differential interference contrast (DIC) images.
17. A non-transitory computer readable storage medium stored with program instructions which, when executed by a computing processor, cause the computing processor to execute a process of segmenting cells in original cell images according to the method of claim 12 .