Biological image transformation using machine-learning models

Inventors

MARIE-NELLY, Herve • VELAYUTHAM, Jeevaa

Assignees

Insitro Inc

Abstract

Described are systems and methods for training a machine-learning model to generate image of biological samples, and systems and methods for generating enhanced images of biological samples. The method for training a machine-learning model to generate images of biological samples may include obtaining a plurality of training images comprising a training image of a first type, and a training image of a second type. The method may also include generating, based on the training image of the first type, a plurality of wavelet coefficients using the machine-learning model; generating, based on the plurality of wavelet coefficients, a synthetic image of the second type; comparing the synthetic image of the second type with the training image of the second type; and updating the machine-learning model based on the comparison.

Core Innovation

The document describes machine-learning model-based generation of enhanced biological microscopy images at scale, including transforming images of biological samples from one modality into synthetic images of other modalities. The transformation is trained to translate bright-field images into synthetic fluorescence, phase, and segmentation modalities, producing enhanced images suitable for downstream evaluation.

The document further describes wavelet-coefficient generation and a Generative Adversarial Network architecture, including a conditional GAN with a generator, discriminator, and PatchGAN. Wavelet-domain representation and losses are used during training, and physics-based image formation modeling may be incorporated to model phase.

In addition, the document discloses hardware-software co-optimization in which illumination patterns are determined using an attention layer and optimized via a spatial light modulator, with a back-propagation module used for the optical and illumination optimization loop. Robustness evaluation is discussed by training classifiers on generated versus real images and discussing batch effects, and treatment evaluation is discussed using enhanced images derived from healthy, untreated, and treated sample sets, compared using biomarker signal identification and distribution- and score-based analysis, with example disease contexts such as NASH and TSC.

Claims Coverage

The independent claims cover a treatment-evaluation workflow that uses a trained machine-learning model to transform images from untreated and treated biological samples affected by a disease of interest, and then compares the transformed images to evaluate the treatment. Across the independent claim set, the main inventive features are the paired image inputs, the transformation via a trained machine-learning model, and the comparative evaluation step.

All independent claims share the same core workflow: receive untreated and treated diseased-sample images, transform each set using a trained machine-learning model, and compare the transformed outputs to evaluate treatment with respect to the disease of interest.

Stated Advantages

Documented Applications