Systematic evaluation of deep learning design choices for image-based organ-on-chip quality assessment

Label-free brightfield microscopy is widely used to monitor organ-on-chip (OOC) cultures, but manual quality assessment can be subjective, labor-intensive, and difficult to standardize across biological conditions. Using a public dataset of 3072 microscopy images from six cell lines, we systematically evaluated four questions relevant to automated OOC quality assessment: (1) how model architecture and image resolution affect performance, (2) whether predictive information is localized or distributed across the image, (3) whether biological metadata provides complementary information beyond the image, and (4) how well learned quality representations transfer across cell lines. We found that (1) both architecture and input resolution influenced performance, with increasing resolution from 224 × 224 to 768 × 768 improving ConvNeXt-Tiny accuracy from 0.852 to 0.910 and improving performance for five of six evaluated convolutional neural networks (CNNs); (2) multiple instance learning and saliency analyses indicated that quality relevant information was distributed across multiple image regions rather than dominated by isolated local defects; (3) biological metadata provided only modest additional benefit, with the best metadata-conditioned model achieving an accuracy of 0.915 compared with 0.910 for image-only classification; and (4) transfer to previously unseen cell lines varied substantially (ROC-AUC 0.783-0.999), while finetuning models with target-cell-line data substantially recovered performance, increasing pooled accuracy from 0.791 to 0.890. Domain-adversarial training did not consistently improve transfer. Rather than introducing a new deep learning architecture, the novelty of this work lies in systematically evaluating these OOC-specific design and generalization factors within a common framework. These findings highlight image resolution, biological-domain validation, and target-domain adaptation as important considerations for robust automated OOC quality assessment.