| The organoid is a three-dimensional cell culture of adult stem cells cultured in vitro.It is highly similar to the source tissue or organ,which can be cultured in vitro on a large scale and conduct high-throughput drug screening.It has a broad application prospect in regenerative medicine,disease research,drug development,etc.Traditional biochemical methods are challenging to analyze individual organoids and cannot continuously observe the growth and development process.Microscopic imaging is a non-destructive testing method that can observe samples repeatedly or continuously,enabling analysis down to individual organoids.Therefore,organoids’ dynamic quantitative analysis method based on microscopic images have essential research value.In response to the urgent needs of drug experimentation and clinical precision treatment,this study focuses on two interdisciplinary scientific issues:organoid detection and tracking of organoid viability evaluation by microscopic image.The research findings include:(1)We propose a model for the dynamic tracking of organoids in high-throughput microscopy imaging.This method decomposes the organoid tracking problem into an organoid detection and matching problem.We propose an organoid detection mode based on the SSD framework and achieve 80.9%detection mAP.In addition,we propose an organoid matching model to realize organoid matching between frames based on multitask learning,which reaches 78.5%matching accuracy.(2)We propose an organoid viability evaluation method as an alternative to manual imaging analysis.The method introduces contrastive learning to mine weak label information of experts for organoid viability evaluation and a multi-head classifier to reduce the dependence on accurate annotation.At the same time,we propose an entropy minimization loss to improve the model’s generalization by introducing sample similarity.This method effectively copes with the labelling divergence among experts and achieves an evaluation performance comparable to that of experts.(3)We propose a method for organoid biomarker estimation based on high-throughput microscopy images.Given the influence of various interferents in the image,this paper introduces multi-instance learning to estimate the biomarker of organoids.To achieve a wide range of biomarker regression,we propose an encoding loss to improve the model convergence speed and training stability.The average relative error of prediction of this method reaches 2.5%,and it is the first time to achieve a non-invasive,high-throughput,whole-process organoid activity evaluation in the way of general biochemical index description.The research results of this paper have promoted the development of the interdisciplinary field of organoid microscopic image analysis and provided a digital morphological evaluation theoretical method for the scientific research and engineering application of organoids,which has scientific significance and practical value. |
