3D Anisotropic Asymmetric Pyramid (AAP-Net) Network Model for Brain Metastases Segmentation on Multi-Sequence MRI

Brain metastases (BM) present a formidable challenge in clinical oncology, demanding precise segmentation and prediction to optimize patient care. Definite detection of BM is crucial for effective treatment planning and prognosis, particularly in radiation therapy. Manual detection of BMs is laborious and prone to error due to their small sizes, irregular shapes, and low contrast in the imaging. The emergence of deep learning technologies has shown promise in automating BM detection, thereby addressing the limitations of manual methods. However, most existing approaches utilize 2D images, overlooking spatial information and employing uniform isotropic images, which deviates from real-life conditions where anisotropic images are more common. This paper presents a novel 3D anisotropic aysmmetric pyramid (AAP-Net) model that integrates advanced deep learning techniques to accurately detect and segment BMs in 3D magnetic resonance imaging (MRI) volumes, capturing the intricate details and spatial relationships crucial for precise diagnosis. By preserving the anisotropic nature of medical images, our model better reflects real-life scenarios, enhancing its efficacy in clinical settings. Through extensive evaluation, we demonstrate the superior performance of our 3D model compared to previous state-of-the-art models. This superiority is evidenced by achieving a mean Dice Similarity Coefficient (DSC) score across K-fold of $0.88~\pm ~0.01$ , Intersection Over Union (IOU) score of $0.80~\pm ~0.02$ , recall and precision values of $0.91~\pm ~0.01$ and $0.93~\pm ~0.01$ respectively using T1 post-contrast MRI images, underscoring the model’s enhanced accuracy and efficacy in BM detection and segmentation.