ENH, MAINT: perform finetuning/hyperparameter optimization for opencv and sam2 detection modules

[adamwitmer]

For open PR's #2 and #3, the lack of fine-tuning and/or systematic hyperparameter optimization for the opencv SimpleBlobDetector and SAM2 models/parameters presents a major limitation of the current work. These procedures should be performed to guarantee optimal performance of the respective segmentation algorithms, ideally in relation to some ground-truth image(s) by using i.e. intersection-over-union (IoU) aka Jaccard score (https://scikit-learn.org/stable/modules/generated/sklearn.metrics.jaccard_score.html) as a metric for quantifying segmentation performance.

For SimpleBlobDetector, the model parameters are described here: https://opencv.org/blog/blob-detection-using-opencv/, and are currently determined using manual tuning based on visual inspection. Similarly, the current implementation of SAM2 algorithm uses pre-trained weights and does not implement any fine-tuning of the model on the "raw" microscopy image data or systematic hyperparameter optimization for the SAM2AutomaticMaskGenerator parameters as described in the source code: https://github.com/facebookresearch/sam2/blob/2b90b9f5ceec907a1c18123530e92e794ad901a4/sam2/automatic_mask_generator.py#L36

[adamwitmer]

As mentioned in the discussion at #3 (comment), the iteration times for performing bubble detection with SAM2AutomaticMaskGenerator are extremely slow (noted as ~6h on ARM M2 Macbook using mps backend and ~20 min per image on Nvidia RTX 4090) and careful consideration should be made when performing hyperparameter optimization with regards to the performance of the model (as measured by ROC-AUC) in relation to the mask settings used for SAM2 detection.

In discussing with Mihee on 2/24/26, she noted that some of the bubbles observed using confocal microscopy (with a 60x) objective are as small as 1 micron in diameter, and the TECAN plate readers maximum objective strength is 10x, such that images taken on TECAN containing very small bubbles require detection of fine-grained detail. One approach may be: sample images with extremely small bubbles in them (labeled as two phases), then determine the minimal SAM2 mask settings that allow detection of the smallest bubble in an image.