Evaluation of Chest CT-Scan Anomaly Detection Models Using Principal Component Analysis

Machine learning-based improvements in anomaly detection, visualization, and segmentation are made possible by the growing digitization of medical imaging, which reduces the workload for medical specialists. Nevertheless, trustworthy labeled data is necessary for supervised machine learning and is sometimes expensive, time-consuming, or impossible to get. Consequently, techniques that need either little labeling (unsupervised methods) or partial labeling (semi-supervised methods) have been used increasingly been used. One approach that can be used to tackle medical imaging tasks like segmentation and classification is anomaly detection, which makes use of both unsupervised and semi-supervised methods. This study investigates the anomaly detection capabilities of deep learning and machine learning models trained on chest CT scans. To adopt this approach, firstly, by evaluating several feature extraction techniques (Principal Component Analysis (PCA), Autoencoder, and variance threshold), PCA achieves the best results on the selected dataset. After that, different machine learning and deep learning models (Isolation Forest, K-means, feed-forward neural networks, and convolutional neural networks) are assessed to detect anomalies using PCA for dimensionality reduction. The experiments show that the NN-Feedforward model with PCA outperformed the other models, also attains remarkably low false alarm rate, demonstrating its effectiveness in reducing false positives and precisely detecting anomalies.