Accurate Anomaly Detection in Medical Images using Transfer Learning and Data Optimization: MRI and CT as Case Studies

Today, medical imaging techniques are widely used to detect a variety of human conditions and diseases. To speed up the diagnostic process, systems are often automated using deep learning methods, which have been proven to yield outstanding results. However, these models may be limited to a specific domain, prone to overfitting, and difficult to update. In this paper, we present a transfer learning approach that achieves high accuracy for detecting anomalies in both MRI and CT images. Our system utilizes the ConvNeXt network, inspired by the Swin Transformer architecture, as a high-level feature extractor from processed and augmented medical images. The resulting data is then optimized using the Edited Nearest Neighbors algorithm, which selects the most effective observations and discards the rest. Finally, the data is passed to the K-Nearest Neighbors classifier to detect anomalies in numerically represented images. Our model achieves high performance for two different types of medical images and can easily be updated without the need for retraining.