Detection of Osseous Metastasis from Bone Scintigrams Using a Combined Global and Local Patch-Based Deep Learning Model

• Main Authors: Lumba, Swailem Neil Angelo, Evangelista, Emmanuel Linus, Martin, Kyla Sydney, Alampay, Raphael, Abu, Patricia Angela R
• Format: text
• Published: Archīum Ateneo 2024
• Subjects: bone scintigraphy; convolutional neural network (CNN); early fusion; global CNN; late fusion; local CNNs; neural networks; osseous metastasis; patch-based analysis; Analytical, Diagnostic and Therapeutic Techniques and Equipment; Biomedical; Computer Engineering; Computer Sciences; Electrical and Computer Engineering; Engineering; Medicine and Health Sciences; Physical Sciences and Mathematics
• Online Access: https://archium.ateneo.edu/intelligent-visual-env/3; https://doi.org/10.1049/ipr2.13311
• Institution: Ateneo De Manila University

Osseous metastasis, or bone metastasis, refers to the spread of cancer cells from their primary site to the bones, often considered by professionals as an indication that cancer has advanced to a level in which it can no longer be cured. As such, early detection and proper treatment both help in improving the patient's quality of life. To aid in the detection of bone metastasis, a Convolutional Neural Network (CNN) has been developed and trained to detect bone metastases from the results of bone scintigraphy. Whole body bone scans were used to train the Global CNN, showing a specificity score of 82.06% and a sensitivity score of 84.56%. Ten Local CNNs-which utilizes the same neural network architecture but uses patches of the original bone scan-were also trained and then merged into the Global CNN using Early Fusion and Late Fusion. Early Fusion improved the sensitivity score to 86.85% while Late Fusion improved the specificity score to 91.41%. These newly developed models were also compared to pretrained models VGG16, ResNet50, and DenseNet121 for further analysis. The model was also compared with a basic vision transformer model in order to gauge performance.