NOA: a versatile, extensible tool for AI-based organoid analysis
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "NOA: a versatile, extensible tool for AI-based organoid analysis".
Title and authors: Tom: So we're looking at this paper called "NOA: a versatile, extensible tool for AI-based organoid analysis," and what it does is put together a general purpose graphical user interface to simplify all that complex AI work with organoid microscopy images.
Jane: It focuses on making the whole process accessible, so instead of juggling ten different tools for detection or segmentation, they’ve built one central interface where you can manage everything from start to finish.
Lu: The authors are from institutions like Helmholtz Munich and TUM, which suggests a solid foundation in both computational science and biological research is behind this development.
Meng: I wonder how robust this GUI is when you try to run it on really large datasets, because the paper mentions caching previous results for detection models and timelapses.
Lalam: The paper emphasizes that NOA integrates modules for detection, segmentation, tracking, feature extraction, custom feature annotation and ML-based feature prediction all into a single workflow.
Tom: That's the big promise; it takes the analysis pipeline—detection through classification—and makes it something you can interact with visually without needing deep programming knowledge.
The paper's summary: Jane: To summarize what this paper is about, they introduce NOA, which is this general purpose graphical user interface designed to handle almost every step of analyzing organoid microscopy images using AI.
Tom: They show how it works by integrating detection, manual adjustment and filtering, segmentation, feature extraction, annotation and prediction into one cohesive workflow shown in their Figure one <ref:2511.01549#pg1>.
Lu: The methodology involves using four pre-trained models for detection—Faster R-CNN, YOLOv3, SSD and RTMDet—and then using the detected bounding boxes to prompt the Segment Anything Model or SAM for instance segmentation masks.
Meng: They also mention that geometric features like area, perimeter and roundness are automatically calculated from those segmentation masks during the segmentation stage so you don't have to calculate that by hand.
Lalam: Beyond detection, they implement common machine learning methods like Random Forest, AdaBoost, MLP and Gaussian Process for automatically predicting custom features based on data currently opened or data in CSV files.
Tom: It seems like the real power is how it allows users to train ML models for feature prediction directly within the interface using a standardized interface they've built.
The paper's improvements: Jane: Now, let’s talk about what they claim is new or improved with NOA; they are highlighting its versatility through three case studies that show it works across different kinds of analysis.
Lu: They demonstrated morphological quantification by using the tool to compute geometric features such as size and roundness during differentiation stages like spheroid I, spheroid II, budding and enteroid.
Tom: And they also showed phototoxicity assessment where they compared phototoxicity measurements from brightfield, GFP and cleaved Caspase-three stained images at zero hours versus fourteen hours <ref:2511.01549#pg1>.
Meng: They used NOA to compare the mean GFP intensity decrease with the mean Caspase-three dye intensity increase substantially, which reflects a progressive loss of viability over time <ref:2511.01549#pg1>.
Lalam: For prediction, they trained ML models using NOA to predict differentiation stages and live dead class labels under TNF alpha induced stress.
Tom: They found that their MLP model achieved the highest accuracy at sixty-eight point fifty-six percent for differentiation stage prediction, while the Random Forest achieved the best result for viability prediction with an accuracy of seventy-two point twelve percent.
Conclusion: Jane: So to wrap up, this paper introduces NOA as a significant advancement because it successfully integrates detection, segmentation, tracking and ML-based prediction into one accessible GUI platform.
Tom: It really bridges that gap between fully automated approaches and the need for deep domain expertise by combining automation with manual refinement capabilities.
Lu: The tool is open-source and available at github dot Meleray organoid analyzer which means researchers can extend it or adapt modules to specific experimental requirements easily.
Meng: If you're deploying this at scale, I'd be looking at how you handle the storage of those detection IDs and cached feature extraction models across different sessions to make sure the system doesn't bog down.
Lalam: I’m quite confident in the results they show, as they provide concrete accuracy numbers for both differentiation stage prediction and viability classification under stress conditions.
Tom: So, NOA is a very versatile tool that lets biologists perform complex AI analysis without needing to be a programmer or spend all their time writing custom code.
Mikhail Konov, Lion J. Gleiter, Khoa Co, Monica Yabal, Tingying Peng
Helmholtz Munich Institute of Molecular Immunology School of Medicine and Health TUM School of Computation Information and Technology TUM
cs.CV
Submitted: 2025-11-03
Updated: 2025-11-03
DOI: 10.1109/ISBI61048.2026.11515532
Code: https://github.com/Meleray/napari-organoid-analyzer
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 80/100
The gist: AI tools can greatly enhance the analysis of organoid microscopy images, from detection and segmentation to feature extraction and classification This work introduces the Napari Organoid Analyzer…
Key concepts
- Organoid Analysis Pipeline
- This refers to the entire sequence of steps taken to analyze an organoid image using artificial intelligence. It includes finding the organoids (detection), separating them into individual parts (segmentation), following them across multiple images (tracking), calculating their properties (feature extraction), and making predictions about their state or type.
- Napari Organoid Analyzer (NOA)
- NOA is a user-friendly, open-source graphical interface built on the napari platform. Its main goal is to combine various state-of-the-art AI algorithms into a single tool, making advanced organoid analysis accessible to researchers who may not have extensive programming knowledge.
- Instance Segmentation
- This is a technique that identifies and outlines every individual object of interest within an image. In the context of NOA, it uses the Segment Anything Model (SAM) prompted by AI detections to create precise masks around each organoid, allowing for accurate measurement and feature calculation for each specific structure.
- ML-based Feature Prediction
- This capability allows users to train machine learning models (like MLP or Random Forest) directly within the tool. These models can then predict complex outcomes, such as the differentiation stage of an organoid or whether it is live or dead, based on its visual features.
Terminology
Summary
AI tools can greatly enhance the analysis of organoid microscopy images, from detection and segmentation to feature extraction and classification This work introduces the Napari Organoid Analyzer (NOA), a general purpose graphical user interface to simplify AI-based organoid analysis
Introduction and Motivation
Organoids are three-dimensional, stem cell–derived tissue models that recapitulate key structural and functional features of human organs AI has the potential to automate almost every step of the analysis pipeline, including detection, segmentation, feature extraction or other supervised prediction A major barrier is the lack of user-friendly implementations, as most algorithms do not provide a graphical user interface (GUI) and hence require substantial programming knowledge Existing tools are typically designed for specific analysis workflows or organoid types rather than general applicability
NOA Architecture and Capabilities
NOA integrates modules for detection, segmentation, tracking, feature extraction, custom feature annotation and ML-based feature prediction into a single GUI The tool interfaces multiple state-of-the-art algorithms and is implemented as an open-source napari plugin for maximal flexibility and extensibility The workflow demonstrates detection, manual adjustment & filtering, segmentation, feature extraction, annotation, and prediction Detection utilizes four pre-trained models from a recent study on AI-based organoid detection [7], based on the Faster R-CNN, YOLOv3, SSD, and RTMDet implementation in mmdetection [11] Segmentation is performed by prompting SAM with the detected bounding boxes to obtain instance segmentation masks Geometric features such as area, perimeter, and roundness are automatically calculated from segmentation masks during the segmentation stage
Experimental Validation
The versatility of NOA is demonstrated through three case studies involving mouse intestinal organoids (mSIOs)
(Morphological Quantification)
NOA was used to compute geometric features such as size and roundness during differentiation stages like spheroid I, spheroid II, budding, and enteroid The tool's ruler annotation function was used to measure wall thickness for specific organoid types
(Phototoxicity Assessment)
NOA was used to compare phototoxicity measurements from brightfield, GFP, and cleaved Caspase-3 (cl.Casp3)-stained images at 0 hours and 14 hours Mean GFP intensity decreases while mean cl. p. 3Casp3-dye intensity increases substantially, reflecting progressive loss of viability
(Prediction)
ML models trained with NOA were used to predict differentiation stages and live/dead class labels under TNF-α-induced stress Table 1 shows that MLP achieves the highest accuracy of 68.56% for differentiation stage prediction, while Random Forest achieves best result for viability prediction with an accuracy of 72. p. 312%
Conclusion and Future Work
NOA significantly advances automated organoid analysis by integrating detection, segmentation, tracking, feature extraction and ML-based prediction into an accessible GUI platform The tool bridges the gap between fully automated approaches and the need for domain expertise in organoid analysis through combining automation with manual refinement capabilities In follow-up work, researchers plan to further streamline the integration of new models by modularizing and standardizing the interface for detection, segmentation, and tracking The tool is open-source and available at https://github.com/Meleray/napari-organoid-analyzer
Acknowledgments
MK acknowledges the financial support from DAAD (German Academic Exchange Service) in the form of a master studies scholarship LG acknowledges support from the Munich School for Data Science (MUDS) This work was supported through the Helmholtz Imaging project ”AIOrganoid” KC and MY are supported by the BMBF 16LW0432, DFG/TRR353-471011418, DFG/CRC1371-395357507 and the DFG-504986538 grants
References
[2] Jonathan M. Matthews, Brooke Schuster, Sara Saheb Kashaf, Ping Liu, Rakefet Ben-Yishay, Dana IshayRonen, Evgeny Izumchenko, Le Shen, Christopher R. p. 4 Weber, Margaret Bielski, Sonia S. p. 4 Kupfer, Mustafa Bilgic, Andrey Rzhetsky and Savas¸ Tay
[3] T. Park, T. K. Kim, Y. D. Han, K. Kim, H. Kim and H. S. Kim
[4] Eva Domenech-Moreno, Anders Brandt, Toni T Lem-metyinen, Linnea Wartiovaara, Tomi P Makel ¨ a and Saara Ollila
[5] Juliet W. Lefferts, Suzanne Kroes, Matthew B. p. 4 Smith, Paul J. p. 4 Niemoller, Natascha D A Nieuwenhuijze, Heleen N Sonneveld van Kooten, Cornelis K van der Ent Jeffrey M Beekman and Sam F B van Beuningen
[6] Timothy Kassis, Victor Hernandez-Gordillo, Ronit Langer and Linda G Griffith
[7] Christina Bukas, Harshavardhan Subramanian, Fenja See, Carina Steinchen, Ivan Ezhov Gowtham Boosarpu Sara Asgharpour Gerald Burgstaller Mareike Lehmann Florian Kofler and Marie Piraud
[8] Giovanni Cicceri, Sebastiano Di Bella, Simone Di Franco, Giorgio Stassi Matilde Todaro and Salvatore Vitabile
[9] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao Chloe Rolland Laura Gustafson Tete Xiao Spencer Whitehead Alexander C Berg Wan-Yen Lo Piotr Dollar and Ross Girshick
[10] napari contributors napari a multi-dimensional image viewer for python doi:10.5281/zenodo. p. 4 p. 43555620
[11] K Chen, J Wang, J Pang, Y Cao Y Xiong X Li S Sun W Feng Z Liu J Xu Z Zhang D Cheng C Zhu T Cheng Q Zhao B Li X Lu R Zhu Y Wu J Dai J Wang J Shi W Ouyang C C Loy and D Lin MMDetection Open mmlab detection toolbox and benchmark arXiv preprint arXiv:1906.07155
[12] J Ma, R Xie, S Ayyadhury C Ge A Gupta R Gupta S Gu Y Zhang G Lee J Kim W Lou H Li E Upschulte T Dickscheid J G de Almeida Y Wang L Han X Yang M Labagnara V Gligorovski M Scheder S J Rahi C Kempster A Pollitt L Espinosa T Mignot J M Middeke J-N Eckardt W Li Z Li X Cai B Bai N F Greenwald D Van Valen E Weisbart B A Cimini T Cheung O Bruck G D Bader and B Wang The multimodality cell segmentation challenge toward universal solutions Nature Methods vol 21 no 6 pp 1103
[13] Nicolas Carion, Francisco Massa, Gabriel Synnaeve Nicolas Usunier Alexander Kirillov and Sergey Zagoruyko End-to-end object detection with transformers in European Conference on Computer Vision 2020 pp 213–229
[14] Daniel B Allan, Thomas Caswell, Nathan C Keim Casper M van der Wel Ruben W Verweij Trackpy v0.5.0 doi:10.5281/zenodo.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
General purpose graphical user interface for AI-based organoid analysis enables biologists without programming experience to perform detection, segmentation, tracking, feature extraction, custom feature annotation and ML-based feature prediction through a single tool.
-
NOA automates analysis using state-of-the-art AI methods by incorporating advanced algorithms for organoid detection and leveraging the foundational Segment Anything Model (SAM) for segmentation.
-
The system allows users to annotate organoids with custom classes, length measurements, counts, or free-form text and train ML models for feature prediction directly within the interface.
-
NOA provides four pre-trained detection models (Faster R-CNN, YOLOv3, SSD, and RTMDet) plus a custom model utilizing SAM features with a DETR-based detection head.
-
Detection supports multi-scale inference with configurable window sizes and downsampling rates, while tracking can align unique detection IDs across frames using TrackPy.
-
Users can manually correct detections by deleting, modifying or adding bounding boxes and filter detections via confidence threshold and minimum diameter sliders to refine results without requiring perfect model accuracy.
-
Instance segmentation masks are obtained by prompting SAM with detected bounding boxes, and signal images can be segmented alongside the original image using the same bounding box set as a prompt for SAM.
-
Geometric features (area, perimeter, roundness) and texture/intensitybased features (mean/total signal intensity within instance masks) are automatically computed per organoid for each attached signal image.
-
Users can annotate organoids with five tools: text, number, classes (multi-label), object boxes, and ruler to draw and measure multiple multisegment lines.
-
The system implements common ML methods like KNN, Random Forest, AdaBoost, MLP, Gaussian Process, and SVC for automatic prediction of custom features by training models on data currently opened or in CSV files.
-
All organoid detections, features, and segmentation masks are stored together in a detection ID-indexed internal storage with caching based on an image hash to restore the internal storage and trained feature extraction models from previous analysis sessions.
Sources
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs