5th International Conference on Gynecology and Obstetrics

Vladislav A. Odintsov has completed his PHD at the age of 33 at the Federal State Budgetary Educational Institution of Higher Education, Northern State Medical University of the Ministry of Health of the Russian Federation. He is the head of Odintsov Clinic. He has over 150 publications that have been cited over 50 times, and his publication h-index is 5. He is a key opinion leader in Russia in the field of minimally invasive breast interventions.

Background: Most practicing physicians consider the BI-RADS system to be subjective and with low agreement between radiologists. Several modifications to the BI-RADS classification have been proposed to improve the consistency of assessment by physicians using models. Ultrasound imaging is an operator-dependent technique, so selecting images to form a dataset for training AI systems is a very challenging task compared to static image acquisition methods.

Objectives: This study assessed the agreement of opinion among 49 radiologists using the BI-RADS classification system and eight breast cyst models proposed by the authors. Based on these cyst models, we developed a method for generating a training dataset from breast ultrasound video recordings, incorporating automated segmentation and quality control for artificial intelligence training.

Methods: We evaluated the interpretations made by 49 radiologists (with 1 to 15 years of experience) who analyzed ultrasound video recordings of 10 verified cases of breast cysts and mixed cystic-solid masses corresponding to BI-RADS categories 2–5. Agreement in category assignment was assessed using Fleiss's method. A three-stage data preparation process was developed: video preprocessing (extraction of regions of interest—ROI), fine-tuning of YOLOv8n-seg and YOLOv11n-seg segmentation models on the open BUSI dataset (798 images), and automated frame extraction based on three different strategies.

Results: The lowest agreement was observed for BI-RADS categories 3 and 4. The use of ultrasound models of breast lesions demonstrated higher effectiveness (0.44 vs. 0.12) and agreement (4-fold). In the analysis of ultrasound video recordings, Strategy 2 provided the largest volume of data (209 frames), whereas Strategy 3 produced a smaller volume but achieved high segmentation agreement (IoU = 0.753).

Conclusion: The study results indicate a low level of inter-radiologist agreement when using the BI-RADS system. Radiologist consensus within a model-based system yields the best results for making accurate clinical decisions. The developed method of segmentation and quality control increases partial automation of the process of training artificial intelligence systems and preparing the necessary data.