Poster #55: Tess Cherlin PhD

Evaluating “Out-of-the-Box” Foundation Vision Models in PMOS Follicle Detection

"Cherlin T (1), Strydesky S (1), Mohammed S (1), Kapoor A (1), Verma SS (1) 1.

Polyendocrine Metabolic Ovary Syndrome (PMOS) is one of the most common endocrine disorders affecting nearly 7-10% of reproductive-aged women worldwide and believed to be underdiagnosed by up to 30%. PMOS is diagnosed by either oligo-anovulation and hyperandrogenism or by the detection of ≥ 20 follicles in at least one ovary when assessed by transvaginal ultrasound (TVUS). AI-driven ultrasound imaging analysis has already been shown to improve diagnostic accuracy in PMOS especially when combined with clinical features. However, there is still no gold standard vision model for PCOM and developing new models can require thousands of images, which is often not feasible in biomedical research with small sample sizes or limited resources. We wanted to see if MedSAM, a foundation model trained on over 1 million medical images, could be fine-tuned for follicle detection in TVUS using a small sample-size. We sourced ovarian TVUS images from Kaggle, sub-setting 348 for training pairs (238 multi-follicular, 110 healthy) and 40 for manual annotation test images (20 per class). MedSAM was fine-tuned using low-rank adaptation (LoRA) with a confidence threshold of 0.89. Performance was evaluated against base MedSAM using Dice score, follicle count MAE, and within 1 follicle. We observed that Dice scores improved by 262% (base MedSAM = 0.147, fine-tuned = 0.530) while follicle count MAE reduced by 70% (Base MedSAM = 7.70, fine-tuned = 2.33). Additionally, the fine-tuned model did a much better job predicting the number of follicles in an ovary. Compared to base MedSAM, the fine-tuned model improved by 32.5 percentage points (base MedSAM = 25%, fine-tuned = 57.5. Using only 348 training and 40 test TVUS images, fine-tuning MedSAM improved follicle detection accuracy over the base model, suggesting that adapted foundation models could be used for accessible PMOS detection methods. These findings are preliminary as the images lacked clinical measurements, menstrual cycle data, and confirmed PMOS diagnoses, limiting interpretability, generalizability, and statistical power. Future work will validate these findings in clinically diagnosed cohorts, expand image diversity, incorporate health record data, explore ultrasound-specific foundation models, and establish a radiologist-annotated ground truth dataset.