Models and datasets

DeLTA comes with pre-trained models, training datasets (that were used to train these models), and demonstration movies.

Models

DeLTA has two modes of operation, selectable with the Config object: “2D” (confluent cells, e.g. agar pads), and “mothermachine”.

The 2D mode uses two deep learning models sequentially: one for segmentation, and one for tracking. The mothermachine mode adds one initial model, that it uses to detect mothermachine chambers.

DeLTA automatically downloads these models as needed. However, you can also use the download_model() function to download them manually (see assets for details such as download location).

  • ROI detection: UNET (5 levels, 64*2^n filters per level), input size: (512, 512, 1)

  • Segmentation: UNET (5 levels, 64*2^n filters per level), input size: (256, 32, 1)

  • Tracking: UNET (5 levels, 64*2^n filters per level), input size: (256, 32, 4)

  • Segmentation: UNET (5 levels, 64*2^n filters per level), input size: (512, 512, 1)

  • Tracking: UNET (5 levels, 64*2^n filters per level), input size: (512, 512, 4)

Training datasets

The training datasets for these models are public and you are free to add them to your own datasets, in order to train more reliable models.

They can be downloaded with the download_training_set() function.

  • ROI detection: 886 samples, 226.90 MiB

  • Segmentation: 2 datasets: train (3294 samples, 45.79 MiB) and train_multisets (9024 samples, 193.96 MiB)

  • Tracking: 2 datasets: train (4055 samples, 127.57 MiB) and train_multisets (7706 samples, 360.28 MiB)

For segmentation and tracking, the train datasets contain only images from the Dunlop lab, while the train_multisets datasets additionally contain images sent to us by the following persons:

  • Simon van Vliet at University of Basel

  • Zoran Marinković and Marianne Grognot in Philippe Nghe’s group at ESPCI Paris

  • Noah Olsman and Daniel Eaton in Johan Paulsson’s group at Harvard

  • Shuai Yang in Fan Jin’s group at Shenzhen Institutes of Advanced Technology

  • Jordi van Gestel and Noam Golan in Avigdor Eldar’s group at Tel-Aviv University

  • Segmentation: 297 samples, 140.92 MiB

  • Tracking: 23655 samples, 2.27 GiB

Demonstration movies

We also provide demonstration movies, such that you can test DeLTA on our data even if you don’t have your own yet.

They can be downloaded with the download_demo_movie() function.

  • 3 positions, 2 channels, 193 time frames: 1160 png images (557.83 MiB)

  • 1 position, 1 channel, 74 time frames: 74 png images (79.65 MiB)