Brain metastases in lung cancers and diagnostic reporting with machine learning methods

Brain metastases (BM) are among the most common complications of lung cancer. Accurate and early detection of these lesions on MRG is essential in diagnosis, treatment planning, and prognosis. However, manual segmentation of metastases is time-consuming, prone to inter-observer variability, and often limited by subtle imaging characteristics. In this thesis, a deep learning-based tumor recognition approach for detecting BM originating from lung cancer using T1c and T2-weighted MRG images is proposed. The CNN model is trained and evaluated on annotated datasets to perform automatic tumor detection. Due to computational limitations, 4D NIfTI volumes were preprocessed into 2D slices, resized to 128×128 pixels, and normalized. After preprocessing and balancing the data, a total of 2.275 2D images were used. The CNN model consists of 17 layers and 3 blocks for binary classification. The occlusion sensitivity and class activation mapping are applied to visualize the model's decision-making process and enhance interpretability. The trained model achieved 94.2% accuracy, 94% sensitivity, 92.83% specificity, 95% F1-score, and an AUC of 98.65% on the test set. While the model shows promising results in many cases, certain cases with low dice scores are analyzed and discussed. This work aims to support the detection and assessment of BMs and to support the development of reliable deep learning solutions in medical imaging field.