anki lab Resources

Resources

Datasets, code and tools from the lab.

HANCOCK — multimodal head and neck cancer data

HANCOCK is a real-world, monocentric dataset of 763 patients with head and neck cancer, assembled with clinical partners at Universitätsklinikum Erlangen. It brings together complementary clinical modalities, including histopathology, medical imaging, blood-based immunophenotyping and structured clinical information. The dataset is intended for multimodal machine learning and precision-oncology research, with the accompanying documentation describing the cohort, data structure and access conditions.

BAGLS — benchmark for automatic glottis segmentation

BAGLS contains 59,250 high-speed endoscopic laryngeal images with manually annotated glottis-segmentation masks. The data were collected across seven institutions and are provided with predefined training and balanced test splits. BAGLS supports reproducible comparison of segmentation methods and makes it possible to train and evaluate models on data from multiple clinical sites.

The recently published BAGLS-VF extension adds pixel-level annotations for the glottal area and the left and right vocal folds to the same videolaryngoscopic frames. This makes the benchmark useful not only for glottis segmentation, but also for multi-structure laryngeal segmentation, glottal-area analysis, laryngeal biomechanics and computer-assisted diagnosis.

DeepD3 — automated dendritic-spine quantification

DeepD3 is an open deep-learning framework for detecting and quantifying dendritic spines in fluorescence microscopy. It predicts spine-probability maps, separates individual spines and derives quantitative morphology from dendrite image data. The released benchmark data include dendrite tracings, spine annotations from multiple human experts and examples spanning different experimental conditions, enabling both method development and independent evaluation.

The SpokeN-100 dataset

SpokeN-100 is an openly available, entirely AI-generated benchmarking dataset for speech recognition and TinyML. It contains the spoken numbers from 0 to 99 generated with 32 different voices in four languages — English, German, French and Mandarin — resulting in 12,800 audio samples. The dataset is designed for benchmarking tasks such as spoken-number and language classification, including models targeting resource-constrained microcontrollers.

The dataset is openly available through Zenodo, while analysis code and cross-validation splits are provided on GitHub. Further details can be found in the original publication (Groh et al., tinyML Research Symposium 2024).

Connectomics game

An interactive introduction to connectomics: explore how neurons and their connections form a network, and learn the basic ideas behind tracing and circuit reconstruction through play. The game is available on our website. The source code is hosted with the anki lab GitHub repositories. A short demonstration video is available on LinkedIn.

ImagesAreExcelSheets

A playful notebook that converts an image into an Excel workbook. Each pixel is represented through spreadsheet cell formatting, with separate RGB, red, green, blue and luminance views. It is a compact way to demonstrate how images can be represented as structured numerical data — and how much visual information can be reconstructed from a table. A short walkthrough video is available on here.