Evaluation of Deep Learning-Based Auto-Segmentation Workflow for Pre-Treated Brain Metastasis

Summary: Accurate detection of small brain metastatic lesions (1-2 mm in size) is critical for patient prognosis. Manual assessment, particularly when dealing with multiple small lesions, is laborious and time-consuming. This study aims to evaluate the effectiveness of several deep learning-based auto-segmentation methods for brain metastasis cases. We utilized a dataset comprising 120 patient cases with manually segmented ground truth labels, focusing on T1 post-contrast MRI scans. Three auto-segmentation methods—NeuralRad, INTContour, and a locally trained model—were evaluated. The performance metrics included sensitivity, Dice scores, and 95% Hausdorff distances for individual segmented lesions. The performance evaluation of the auto-segmentation algorithms indicates significant advancements in detecting and segmenting metastatic lesions. Initial results demonstrate that NeuralRad achieves high Dice scores (0.89) for larger lesions (>39mm3) and shows improving accuracy for smaller lesions (< 39mm3), underscoring the potential of these algorithms for integration into clinical workflows. This study highlights the robustness of automated brain metastasis detection and segmentation using deep learning techniques. The accurate identification of both large and small lesions by these methods has the potential to revolutionize the clinical management of brain metastasis. Ultimately, these advancements are expected to enhance patient care, improve treatment planning, and foster collaborative research in neuro-oncology. Kevin Cao, MD (Presenting Author) - Baylor College of Medicine; Caleb Stewart, MD (Co-Author) - Baylor College of Medicine; Piyush Pathak, MD, MPH (Co-Author) - Baylor College of Medicine; Yiding Han, PhD (Co-Author) - Baylor College of Medicine; Baozhou Sun, PhD (Co-Author) - Baylor College of Medicine; Zaid Siddiqui, MD (Co-Author) - Baylor College of Medicine