Determine which voxel classification method is most effective on the MouseLight data

[shreyasingh1]

Aim to answer the question - which voxel classification method is most effective on the MouseLight data.

This will entail: (1) Reproducing the AUC curves shown in this poster for the Voxel-Based Logistic Classifier and the Multilayer Perceptron. Incorporating and documenting this code in the brainlit repository. Curves will be generated using code similar to this script

(2) Generating an AUC curve for the Flood Filling Networks method (tutorial here) on the benchmarking data. According to the tutorial, FFN doesn't perform well on regions with no cell body so it will not likely perform well on the benchmarking data.

(3) Generating an AUC curve for the APP2 method (tutorial here) on the benchmarking data. According to the tutorial, APP2 appears to struggle with sensitivity especially with large background spaces in images, which are present in the benchmarking data.

(4) Comparing the Voxel-Based LC, MLP, FFN, and APP2 AUC curves to determine which is most effective on the Mouselight Benchmarking data.

[shreyasingh1]

@tathey1 Here are my initial thoughts for this issue. What are your thoughts on producing the AUC curves from the benchmarking data (like your poster) versus producing them from the whole brain data. It seems FFN and APP2 could perform better on cell body images, so perhaps that's something we could explore after the above is done as well?

[tathey1]

I would recommend starting from the benchmarking data first. Beyond just AUCs, you should also set up code that makes it possible to visualize (e.g. in napari, vaa3d) the algorithms' results on image subvolumes.