TRANSFECTION IMAGE FOR CELL WALL ANNOTATIONS

Abstract

Several aspects of the revelation concern techniques for detecting cell structures in microscopic images. In particular, techniques are described for detecting cell walls, i.e., cell margins, in microscopic images, such as phase-contrast images. A machine-learned algorithm, such as an artificial neural network, can be used for this purpose. Techniques are described for generating annotations as the basis for training the machine-learned algorithm, for example, based on the fluorescence channel of transfection image data. Further actions can then be performed based on appropriately localized cell walls, such as determining cell instance annotations that segment cell instances for training another machine-learned algorithm.

Claims

[1] Computer-implemented method, the method comprising: - Obtaining (3005, 3105) transfection image data comprising a fluorescence channel (101) and a non-fluorescence channel (105), wherein the transfection image data depicts a collection of cells, not all of which express a dye, - performing (3010, 3110) an automated image processing of the fluorescence channel (105) to locate boundary regions (112) between fluorescence areas and non-fluorescence areas, - generating (3015, 3115) cell wall annotations (184) along the localized border regions, and - Training (3020, 3120) a machine-learned algorithm (190) based on the cell wall annotations as ground truth and the non-fluorescence channel (105) as input. [2] A computer-implemented method according to any one of the preceding claims, the method further comprising: - generating (3015, 3115) non-cell wall annotations (185) in image areas adjacent to the localized boundary regions (12), wherein the training (3020, 3120) is further based on the non-cell wall annotations as further ground truth. [3] A computer-implemented method according to any one of the preceding claims, the method further comprising: - generating (3015, 3115) unknown annotations in image areas that have a minimum distance to all localized boundary regions (112), wherein the training (3020, 3120) is further based on the unknown annotations as a further ground truth. [4] The computer-implemented method of claim 3, wherein the unknown annotations are further generated based on an estimate of the presence of cells in the corresponding image areas based on the non-fluorescence channel (105). [5] The computer-implemented method of claim 3 or 4, wherein, when training the machine-learned algorithm (190), a prediction of the machine-learned algorithm in regions with the unknown annotation is not included in a loss function. [6] Computer-implemented method according to claim 3 or 4, wherein, during training of the machine-learned algorithm (190), a prediction of the machine-learned algorithm in areas with the unknown annotation is weighted into a loss function, wherein a weighting is determined as a function of a distance to at least one other annotation. [7] A computer-implemented method according to any one of the preceding claims, the method further comprising: - Applying (3025, 3125) the machine-learned algorithm (190) to localize cell walls in a non-fluorescence channel of inference image data. [8] A computer-implemented method according to claim 7, the method further comprising: - based on the localized cell walls in an output of the machine-learned algorithm (190): performing (3030) an instance localization of individual cells in the inference image data. [9] The computer-implemented method of claim 8, wherein performing instance localization detects contiguous regions and then distinguishes between cell instance and background for each contiguous region. [10] A computer-implemented method according to any one of the preceding claims, the method further comprising: - applying (3125) the machine-learned algorithm to localize cell walls in the non-fluorescence channel of the transfection image data or other image data, - based on the detected cell walls: performing (3130) an instance localization of individual cells in the non-fluorescence channel, and - Training (3135) another machine-learned algorithm based on the instance localization as ground truth and the non-fluorescence channel as input. [11] A computer-implemented method according to any one of the preceding claims, further comprising: - Evaluating a spatial structure of fluorescent regions in the fluorescence channel, with the cell wall annotations still being generated based on the spatial structure. [12] The computer-implemented method of any preceding claim, wherein training the machine-learned algorithm (190) uses a consistency loss function or a loss function that minimizes an entropy measure of an output of the machine-learned algorithm (190) in at least some image regions. [13] Data processing device arranged to carry out a method according to one of the preceding claims.

Images