Deepcell

🚀 Features

<details>
<summary>Add support for ISBI inputs to the metrics package msschwartz21 (107)</summary>

This addition to the metrics package makes it possible to directly input ISBI style outputs into our metrics pipeline. This simplifies the process of benchmarking against competitor models which tend to output ISBI style tracks.
</details>


🐛 Bug Fixes

<details>
<summary>Implement correct usage of crop parameter in CellTracker msschwartz21 (108)</summary>

- Corrects an error where the crop parameter was not being set during tracking and inference
- Removes deprecated post processing functions from the CellTracker
</details>


🧰 Maintenance

<details>
<summary>REL: pin numpy and update deepcell-toolbox dependency rossbar (110)</summary>

A patch release to upper-bound numpy to prevent errors related to expired scalar deprecations. Bumps the `deepcell-toolbox` dependency as well, which already has more comprehensive fixes for this issue.
</details>

🐛 Bug Fixes

<details>
<summary>Correct keys in `correct_shifted_divisions` msschwartz21 (105)</summary>

Different key names were used in `correct_shifted_divisions` which caused issues in some downstream analyses in the model-registry. This PR standardizes all keys in the metrics package to avoid confusion.
</details>

🐛 Bug Fixes

<details>
<summary>Classify divisions that are +/- 1 frame as correct msschwartz21 (103)</summary>

Recent reviews of tracking predictions have identified a failure mode in the current metrics packages. Different segmentation predictions can sometimes lead to a cell dividing in one frame before or after the frame assigned to the division in the ground truth. Currently this leads to that division counting as both a false positive and a missed division. This PR introduces a new metrics function that identifies these events and corrects the metrics to classify this division as correct. Additionally a new metrics class for tracking (`TrackingMetrics`) has been introduced to coordinate running all of the necessary metrics functions.

In the current test split, applying the new metrics pipeline led to the following changes in metrics:
| Metric | Old | New |
|-------------------------------|------|------|
| Total divisions | 181 | 181 |
| Correct divisions | 139 | 154 |
| False negative division | 27 | 13 |
| False positive division | 40 | 26 |
| Mismatch division | 15 | 14 |
| Division Recall | 0.84 | 0.92 |
| Division Precision | 0.78 | 0.86 |
| Division F1 | 0.81 | 0.89 |

🧰 Maintenance

<details>
<summary>Bump to 0.6.1 msschwartz21 (99)</summary>


</details>

<details>
<summary>Enable option for fixed size crops in get_image_features vanvalen (98)</summary>

What
Add a flag to get_image_features to allow for doing fixed sized crops rather than crop and resize.

Why
Crop and resize removes information about cell size, which is useful for cell tracking and also learning dynamic representations of cell behavior.
</details>

🚀 Features

<details>
<summary>Update metrics for evaluating tracking performance msschwartz21 (95)</summary>

This PR introduces several substantial changes
- Reorganization of functions with the addition of two new modules: `metrics` and `trk_io`. Backwards compatible imports were maintained whenever possible.
- `load_trks`, `trk_folder_to_trks`, `save_trks`, `save_trk`, `save_track_data` from `utils` to `trk_io`
- `match_nodes`, `contig_tracks` from `isbi_utils` to `utils`
- `classify_divisions`, `calculate_summary_stats` from `isbi_utils` to `metrics`
- `benchmark_division_performance` deprecated in `isbi_utils` and renamed to `benchmark_tracking_performance` in `metrics`
- Fixes bugs in how we built graphs of tracks and compared between ground truth and predictions
- Originally we converted lineage data to isbi format prior to generating a graph. This intermediate step unintentionally removed any discontinuities that were present in a lineage. There is now a new function `deepcell_tracking.utils.trk_to_graph` that faithfully converts lineage data to a graph without any intermediate steps.
- The use of a `node_key` generated by `match_nodes` unintentionally dropped lineages if more than one predicted lineage was mapped onto a single ground truth lineage. Instead of mapping cell ids when we create the graph, we instead map cell ids on the fly when we are comparing graphs which eliminates the risk of accidentally dropping lineages from consideration.
- Introduces Association Accuracy as a new metric that evaluates how many edges in the tracking graph are correctly assigned. This score discounts edges involved with a division, but does detect discontinuities in lineages.
- Introduces Target Effectiveness as a new metric that evaluates how many cells in a lineage are correctly assigned to the lineage.
</details>


🧰 Maintenance

<details>
<summary>Bump version to 0.6.0 msschwartz21 (97)</summary>


</details>

🧰 Maintenance

<details>
<summary>Drop support for python 3.6 and bump deepcell-toolbox requirement msschwartz21 (94)</summary>

Updates deepcell-toolbox to ~=0.11.2.
</details>