Datasets:

---

rajpurkarlab

/

ReXGroundingCT

like 35

Follow

Rajpurkar Lab / Harvard Medical AI 120

Description:
This PR introduces several improvements to rexrank_eval.py to ensure more accurate instance matching and expand compatibility with datasets outside of ReXGroundingCT.

Motivation:
The previous implementation utilized a greedy matching approach for prediction and ground truth (GT) instances, which could lead to sub-optimal pairings. Additionally, it assumed that GT instances were pre-labeled. While this holds true for ReXGroundingCT, other datasets might not have pre-labeled GTs. This PR addresses both limitations.

Summary of Changes:

1. Optimal Instance Matching (Hungarian Algorithm): Replaced the greedy matching approach with the Hungarian matching algorithm (scipy.optimize.linear_sum_assignment). This ensures that the global pairing between GT and prediction components yields a more accurate and higher mean-matched dice score.
2. Optimized Dice Score Computation: Instead of calculating dice scores on the fly during the while-loop, the script now pre-computes and stores the dice score values for overlapping instances in a score matrix prior to assignment. This avoids redundant calculations.
3. Dynamic GT Connected Components: Added a gt_instances_labeled flag to evaluate_finding. If a dataset's GT instances are not pre-labeled (distinct integers), the script will now run connected component filtering (cc_filter) on the ground truth to label them on the fly.

Please let me know if you would like any adjustments to the implementation!