Sample Size Estimation for Chest X-ray Classification with Foundation Models

Dmitry Umerenkov, Viktor Gombolevskiy, Nikolay Nechaev, Evgeniia Przhezdzetskaia, Dmitry V. Dylov

Keywords: Foundation Models, Sample Size Estimation, Learning Curve Analysis, Chest X-ray Classification

TL;DR: We show that the performance of foundation models on chest X-ray classification can be accurately predicted from just 50 labeled samples, enabling researchers to drastically reduce annotation costs.

Abstract: The integration of deep learning models into clinical practice, particularly in radiology, is often hindered by the need for large, meticulously labeled datasets, which entails significant time and financial costs. While foundation models substantially reduce this dependency, a critical question remains: what is the minimum amount of annotated data sufficient to achieve clinically acceptable accuracy? In this work, we introduce a methodology for accurately predicting sample size requirements by modeling learning curves with a power law. Our study demonstrates that modern foundation models, such as XrayCLIP and XraySigLIP, not only outperform traditional architectures but also achieve high ROC-AUC scores with significantly fewer training examples. A key finding of our research is the evidence that the learning dynamics observed with a sample of just 50 labeled cases can predict the model's asymptotic performance with high precision. Thus, our study offers a scientifically grounded approach to optimizing the data annotation process, enabling researchers and clinicians to minimize costs and accelerate the development of reliable diagnostic tools

Primary Area: foundation or frontier models, including LLMs

Submission Number: 24602