One Snapshot, Many Clues: Inverse Protocol Prediction from Single-View Spheroid Images

Prateek Mittal, Ayush Srivastava, Joohi Chauhan

Keywords: Segmentation, Image Reconstruction, Deep Learning, Multi Task Inference, Cell Morphology, Explainability

Abstract: Understanding how experimental protocols shape spheroid morphology is crucial for advancing 3D cell culture research, yet reconstructing these conditions from imaging alone has remained elusive. We used deep learning frameworks which are able to infer the full experimental protocol including cell line, medium, seeding density, timepoint, formation method, microscope, and magnification from a single bright-field spheroid image. Using the SLiMIA dataset of ~8,000 annotated images spanning diverse culture conditions, we cast this as a structured multi-label prediction task and benchmarked a spectrum of models, from CNNs and transformers to hybrid and dependency-aware architectures. Our approach integrates segmentation for morphology extraction, domain-adversarial training, and morphologically informed augmentation to improve robustness across imaging setups. Results show an average accuracy of 95.23% across protocol components, with hybrid models such as CoAtNet excelling in balancing efficiency and accuracy, while feature-augmented and hierarchical models contribute interpretability and consistency. Grad-CAM analyses confirm that predictions rely on biologically meaningful features (e.g., compactness, necrotic core structure), while highlighting dataset-driven artifacts in replicate and magnification tasks.

Primary Area: interpretability and explainable AI

Submission Number: 17627