# Artificial intelligence models and combined scoring approaches for endometrial receptivity assessment in in vitro fertilization

## TL;DR

This study explores how artificial intelligence can help assess endometrial receptivity during IVF by combining image analysis and clinical data to improve pregnancy prediction.

## Contribution

The novel contribution is a composite scoring system integrating AI-based image analysis with clinical metrics to enhance endometrial receptivity assessment.

## Key findings

- The composite score outperformed individual AI components in predicting biochemical pregnancy (AUC 0.94).
- AI models showed improved prediction when combined with clinical probability metrics.
- Prospective multi-center validation is needed for broader clinical application.

## Abstract

This retrospective study evaluates the effectiveness of artificial intelligence (AI) models and integrated scoring systems in assessing endometrial receptivity during in vitro fertilization (IVF). We combine AI-driven image analysis for automated endometrial segmentation (U-Net with VGG16 encoder) and embryo quality classification (VGG16) with a patient-specific clinical probability metric (SART) to generate a composite score. Ultrasound images of the endometrium were processed using a standardized pipeline (denoising, normalization, ROI cropping, augmentation), enabling automated quantification of thickness, echogenicity/pattern, and derived measures. Embryo quality was assessed by a fine-tuned classifier benchmarked against expert embryologists. The composite score is defined as: CS = (EQS + ERS)/2 × PSART/10 where EQS is the AI-derived embryo quality score (0–10), ERS is the AI-derived endometrial receptivity score (0–10), and PSART is the SART pregnancy probability scaled to 0–10. In internal validation, the composite score demonstrated higher discriminative performance for biochemical pregnancy than individual components (AUC 0.94 vs. EQS 0.88 and ERS 0.85). While integrated scoring improved prediction relative to single-source models, generalizability is limited by the single-center, retrospective design and modest dataset size augmented synthetically. Prospective, multi-center validation with live-birth outcomes and incorporation of explainable AI (e.g., saliency/attribution maps) are needed to support clinical deployment. These findings suggest that AI-based models, when embedded in multidimensional scoring frameworks, may help optimize embryo transfer timing and support precision IVF, complementing—not replacing—clinical expertise.