Optimizing Medical Image Classification: Leveraging Advanced Segmentation Models for Enhanced Object Detection in Real-World Scenarios

Abstract

In medical imaging, classifiers typically rely on objects of interest, such as nodules, being centered and fully visible within the frame to effectively learn the underlying relationships necessary for accurate classification. However, in real-world scenarios, such conditions are not always met, leading to challenges in classification accuracy. This paper explores the enhancement of classifier performance by incorporating advanced segmentation models, including SAM, SAM2, and Dinov2, all of which are pretrained for automatic semantic segmentation. We compare the results of integrating these modelsto preprocess and segment medical images, thereby improving the classifier's ability to correctly identify and classify objectsin non-ideal conditions. The study aims to demonstrate how leveraging state-of-the-art segmentation techniques can bridge the gap between idealized assumptions and real-world medical data, ultimately leading to more robust and accurate classificationoutcomes.