A Coherent Approach-Based Fine-Tuning of Segment Anything Model Plus Watershed Algorithm for Instance Segmentation of Mitochondria in Electron Microscopy Images

Increasingly available ultrastructural data from a continuously growing diversity of experimental conditions are driving new opportunities for fruitful neuroscientific hypotheses tested in intracellular compartments such as the nanoscale roles of, e.g., the mitochondria. While state-of-the-art deep CNNs demonstrate advanced capabilities, they can exhibit a degree of fragility; often yielding coarse, high-frequency oscillatory outputs characterized by discontinuities and false positives, even in the context of straightforward segmentation of mitochondria. Traditionally, the leading methodologies in medical image segmentation have utilized variations of the encoder-decoder framework, notably including the U-Net architecture. The SAM does not perform as well, since it has not been explicitly trained for the task and does not demonstrate user-interactive, over one billion annotations mostly for natural images. However, the SAM may be applied to segment anything, including medical image segmentation challenging new datasets. This research tackles the complex task of implementing domain adaptation for task of mitochondria segmentation in electron microscopy images obtained from various tissues and species, employing deep learning methodologies. A thorough analysis is conducted to evaluate the effectiveness of SAM in segmentation tasks in medical imaging and to measure its performance on volumetric datasets derived from electron microscopy, and show that it is powerful at segmenting instances even under challenging imaging conditions. We provide a fine-tuning SAM which can be naturally trained by SAM at an exemplary scale, benefiting from a diverse and large dataset over one million image masks in 11 modalities. This model would be able to perform precise segmentation for a diverse array of targets across different imaging conditions at the level of performance of specialized U-Net models, or even better. A visual comparison is shown between our fine-tuning SAM model and U-Net, along with an examination of different watershed post-processing strategies to discriminate between adjacent or conjoined instances. Results from our experiments show that the method suggested is fast, very robust, and accurate, with an imbibed model that has improved learning capability. A comprehensive sensitivity analysis has been carried out, and an ablation study using most popular metrics segmentation evaluation is performed, the results derived from the Jaccard Index coefficient, Dice Similarity Coefficient, Aggregated Jaccard Index, and Panoptic Quality reinforce the strength and effectiveness of the proposed modules in enhancing the performance of mitochondrial segmentation.