Fix long evaluation time on microscopy images

[mariehbourget]

Checklist

GitHub

• I've given this PR a concise, self-descriptive, and meaningful title
• I've linked relevant issues in the PR body
• I've applied the relevant labels to this PR
• I've assigned a reviewer

PR contents

• I've consulted ivadomed's internal developer documentation to ensure my contribution is in line with any relevant design decisions
• I've added relevant tests for my contribution
• I've updated the relevant documentation for my changes, including argparse descriptions, docstrings, and ReadTheDocs tutorial pages

Description

This PR aims to fix long evaluation time on large microscopy images with the axons and myelin segmentation task.

• It adds a object_detection_metrics param in the evaluation_parameters of the config file.
• The default is True.
• When True, it computes lesions true positive and false detection rates + Hausdorff score at evaluation time.
• When False, those metrics are skipped (they are not appropriate for axon+myelin segmentation and would create the bottleneck when running evaluation on large images).
• When False, the "painted" predictions in the pred_masks are also skipped since they are based on lesions true positive and false detection rates.
• Doc and microscopy tutorial are updated accordingly.

Linked issues

Fixes #975

[coveralls]

Pull Request Test Coverage Report for Build 2783521423

Warning: This coverage report may be inaccurate.

This pull request's base commit is no longer the HEAD commit of its target branch. This means it includes changes from outside the original pull request, including, potentially, unrelated coverage changes.

• For more information on this, see Tracking coverage changes with pull request builds.
• To avoid this issue with future PRs, see these Recommended CI Configurations.
• For a quick fix, rebase this PR at GitHub. Your next report should be accurate.

Details

• 14 of 20 (70.0%) changed or added relevant lines in 3 files are covered.
• 14 unchanged lines in 2 files lost coverage.
• Overall coverage increased (+9.9%) to 79.09%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
ivadomed/evaluation.py	8	14	57.14%

Files with Coverage Reduction	New Missed Lines	%
ivadomed/metrics.py	5	96.09%
ivadomed/main.py	9	55.52%

Totals
Change from base Build 2771735045:	9.9%
Covered Lines:	4883
Relevant Lines:	6174

---

💛 - Coveralls

[mariehbourget]

This PR is now ready for review.
It adds a parameter to skip object detection metrics that were not useful for the axon+myelin tasks and causing a significant bottleneck when evaluating large images (as of #975).
The full description of the PR changes are in the first comment above.

Expected behavior:

With object_detection_metrics=True:

• Hausdorff distance is included in the metrics printed in the terminal

2022-08-12 14:51:29.097 | INFO     | ivadomed.testing:test:88 - {'dice_score': 0.21474601221146625, 'multi_class_dice_score': 0.1535026128012612, 'precision_score': 0.22883224487304688, 'recall_score': 0.2024238745931438, 'specificity_score': 0.7721640585401799, 'intersection_over_union': 0.12044042384784692, 'accuracy_score': 0.629604126767127, 'hausdorff_score': 0.0}

• The evaluation_3Dmetrics.csv file in results_eval has values in columns lfdr_class0 and ltpr_class0.
• "painted" predictions are in pred_masks.

With object_detection_metrics=False

• No Hausdorff distance in the metrics printed in the terminal

2022-08-12 14:51:29.097 | INFO     | ivadomed.testing:test:88 - {'dice_score': 0.21474601221146625, 'multi_class_dice_score': 0.1535026128012612, 'precision_score': 0.22883224487304688, 'recall_score': 0.2024238745931438, 'specificity_score': 0.7721640585401799, 'intersection_over_union': 0.12044042384784692, 'accuracy_score': 0.629604126767127}

• The evaluation_3Dmetrics.csv file in results_eval does not have values in columns lfdr_class0 and ltpr_class0 (the columns are there but are empty).
• No "painted" predictions in pred_masks.

Without any evaluation parameters:

• Same behavior as True.

[dyt811]

Should use dict.get() pattern to avoid invalid key error as upstream JSON is never fully validated.

[LuluSzutu]

since the key 'object_detection_metrics' has already defined in 'config_default.json' file, so it wouldn't be invalid key error show up here.

[LuluSzutu]

Code has been reviewed, and test runed, should be 'good to merge'.
But have one concern about skipping the metrics of true positive and false detection rates. I feel TP and FN rates are the important for evaluate the image segmentation result. Maybe we should let the user has a choice of skipping only the Hausdorff score and skipping the TP FN as well. This can be addressed in a separate issue if need.