Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers, with a median survival of under a year for patients with metastatic disease. Traditional cancer models — such as 2D cell lines or animal experiments — fail to capture the complex architecture and genetic diversity of human tumors, limiting their usefulness for testing new treatments.
Organoids are three-dimensional, lab-grown mini-tumors derived from actual patient tissue. They preserve the genetic mutations, tissue structure, and drug resistance patterns of the original cancer, making them much more realistic models for testing therapies — including immunotherapies. However, testing immunotherapy in organoids requires co-culturing them with immune cells, which is technically difficult.
To address the lack of a standardized imaging dataset, the researchers created OrganoIDNetData — a curated collection of 180 phase-contrast microscopy images containing 33,906 individual organoids. The dataset includes both human-derived and mouse-derived PDAC organoids, all co-cultured with immune cells (PBMCs).
The organoids were grown in Matrigel domes and imaged every 4 hours over 100 hours using an automated live-cell imaging system. This time-lapse approach allowed researchers to track organoid growth, regression, and response to the presence of immune cells in real time.
The dataset is publicly available and designed as a shared benchmark so that different research groups can compare the performance of their organoid segmentation algorithms on the same standardized images — something that was previously impossible because each lab tested on its own private dataset.
Each image in the dataset was manually segmented by two independent expert annotators using Labkit software in Fiji. Where the two annotators disagreed, they consulted and reached a consensus ground truth. This rigorous manual annotation process ensures the AI models learn from high-quality, reliable labels.
Two state-of-the-art AI segmentation algorithms were trained and validated on the dataset: Cellpose, a convolutional neural network designed for cellular segmentation, and StarDist, which uses a star-convex polygon approach to detect individual organoid boundaries. Both were trained for 1,000 epochs and made publicly available as pre-trained models for the community to use.
The training set included 713 images, the validation set 179, and an augmented set of 1,426 images was created by applying data augmentation. Ten full field-of-view images were reserved as a held-out test set to evaluate final model performance on completely unseen data.
Both Cellpose and StarDist performed well on the dataset, but Cellpose showed a slight advantage in reducing false positives — it was less likely to incorrectly identify non-organoid structures as organoids. StarDist tended to over-segment dense regions, identifying too many small structures in images with many overlapping organoids.
Murine (mouse-derived) organoid images were significantly harder to segment than human samples because mouse organoids grow in much denser clusters, leading to higher false negative rates — organoids that AI algorithms missed entirely — for both methods.
The overall inter-annotator agreement between the two human experts was only 21%, reflecting the genuine complexity of the dataset and the difficulty of segmenting organoids co-cultured with immune cells. This complexity makes the dataset a challenging and realistic benchmark for AI development.
A standardized, public organoid imaging dataset with pre-trained AI models dramatically lowers the barrier for labs to adopt automated organoid analysis. Previously, each lab had to build its own segmentation pipeline, creating inconsistent and non-comparable results across studies.
With reliable AI-powered segmentation, researchers can now quantify thousands of organoids automatically — tracking how they shrink or grow in response to different immunotherapy combinations in real time. This accelerates the search for effective PDAC treatments and brings personalized cancer medicine closer to reality.