OrganoidViT: a vision transformer-based deep learning model for segmenting viable and nonviable organoids in bright-field images

The viability assessment of patient-derived tumor organoids is essential for preclinical drug screening, with microscopic imaging serving as a key method for evaluating drug effects. Traditional image analysis methods, such as manual evaluation and fluorescence staining, suffer from low efficiency, dye toxicity, and fluorescence degradation, making them unsuitable for high-throughput drug screening. Additionally, current artificial intelligence (AI)-based tools face challenges in precise segmentation for the accurate quantitative analysis of viable and nonviable organoids. To address these challenges, we introduce OrganoidViT, a novel deep learning model utilizing vision transformer technology for the precise segmentation of viable and nonviable colorectal cancer organoids in bright-field microscopy images, ensuring an accurate efficacy assay of different drugs. Trained on custom datasets of colorectal cancer organoids prepared using microdroplet technology, OrganoidViT facilitates high-throughput segmentation without fluorescence imaging, with experimental results indicating superior performance of OrganoidViT (accuracy of 99.7%) compared to manual annotation and traditional convolutional models. The capability of pixel-level segmentation enables accurate morphological assessments, including area, pellucidity, roundness, and grayscale entropy, which are critical for evaluating growth conditions and the effects of different therapeutic agents on colorectal cancer organoids. Thus, OrganoidViT shows promise for preclinical drug screening, enhancing both the efficiency and accuracy of organoid-based testing.