Systems and methods for effect size optimization of object classification
Methods and systems are provided herein for automatic object classification. In an example, a method includes receiving one or more images of a plate including a plurality of wells, each well including a plurality of cells, the plurality of wells including a first control well, a second control well, and at least one test well, classifying one or more cells from the at least one test well using a trained classification model, the trained classification model trained based on training data including instance images of the first control well and the second control well formed from the one or more images of the plate and further based on an effect size loss function, and outputting an indication of the classified one or more cells.
1 . A method, comprising:
receiving one or more images of a plate including a plurality of wells, each well including a plurality of cells, the plurality of wells including a first control well, a second control well, and at least one test well;
classifying one or more cells from the at least one test well using a trained classification model, the trained classification model trained based on training data including instance images of the first control well and the second control well formed from the one or more images of the plate and further based on an effect size loss function; and
outputting an indication of the classified one or more cells.
2 . The method of claim 1 , wherein classifying the one or more cells from the at least one test well using the trained classification model comprises:
generating, from the one or more images of the plate, a respective test instance image of each cell of a plurality of cells in the at least one test well;
entering each respective test instance image as input to the trained classification model; and
receiving a classification score for each respective test instance image as output from the trained classification model, wherein the classification score for a given respective test instance image indicates a likelihood a cell imaged in the given respective test instance image has an effect that matches an effect of cells in the first control well or the second control well.
3 . The method of claim 2 , wherein outputting the indication comprises outputting a heat map illustrating, for each well of the plate, a respective mean classification score calculated from classification scores output by the classification model.
4 . The method of claim 1 , wherein the trained classification model is trained based on training data including first instance images of cells in the first control well and second instance images of cells in the second control well, the first instance images and the second instance images formed from the one or more images of the plate, the training data further including a first ground truth label for each first instance image and a second ground truth label for each second instance image.
5 . The method of claim 4 , further comprising training an initial classification model to form the trained classification model, the training including:
receiving, from the initial classification model, a respective classification score for each input first instance image and each input second instance image;
determining, for each input first instance image, a respective first loss based on the received respective classification score for the input first instance image relative to the first ground truth label for the input first instance image;
determining, for each input second instance image, a respective second loss based on the received respective classification score for the input second instance image relative to the second ground truth label for the input second instance image;
calculating a cost function based on each first loss, each second loss, and the effect size loss function; and
updating the initial classification model based on the cost function to form the trained classification model.
6 . The method of claim 5 , further comprising calculating the effect size loss function based on the respective classification score for each input first instance image and each input second instance image.
7 . The method of claim 6 , wherein calculating the effect size loss function based on the respective classification score for each input first instance image and each input second instance image comprises:
calculating a first mean classification score and a first standard deviation based on the respective classification score for each input first instance image;
calculating a second mean classification score and a second standard deviation based on the respective classification score for each input second instance image; and
calculating the effect size loss function based on the first mean classification score, the second mean classification score, the first standard deviation, and the second standard deviation.
8 . A system, comprising:
memory storing instructions; and
one or more processors configured to execute the instructions to:
receive one or more images of a plate including a plurality of wells, each well including a plurality of cells, the plurality of wells including a first control well, a second control well, and at least one test well;
classify one or more cells from the at least one test well using a trained classification model, the trained classification model trained based on training data including instance images of the first control well and the second control well formed from the one or more images of the plate and further based on an effect size loss function; and
output a visualization of the classified one or more cells for display on a display device.
9 . The system of claim 8 , wherein the one or more images are fluorescent microscopy images.
10 . The system of claim 8 , wherein the effect size-based loss function is determined based on mean and standard deviation values of batches of output from the classification model during training of the classification model.
11 . The system of claim 10 , wherein each instance image of the first control well includes a first ground truth annotation and each instance image of the second control well includes a second ground truth annotation, each first ground truth annotation indicating that an associated instance image of the first control well belongs to a first class, each second ground truth annotation indicating that an associated instance image of the second control well belongs to a second class, wherein the output from the classification model during training of the classification model includes a regression value for each input instance image indicative of a likelihood that the input instance image belongs to the first class or to the second class.
12 . The system of claim 8 , wherein the classification model is a convolutional neural network.