AI Diagnosis of Orbital Invasion Caused by Malignant Nasal and Paranasal Sinus Tumors Using Transfer Learning

Whether malignant nasal and paranasal sinus tumors have invaded beyond the orbital periosteum is crucial information in determining the surgical approach, specifically whether to preserve or remove the orbital contents. Known imaging findings, such as ‘involvement of the extraocular muscles’ and ‘infiltration into the orbital fat tissue,’ are labeled for each case by head and neck radiology specialists. Using these labeled data, a convolutional neural network (CNN) model was trained with transfer learning to determine the presence of invasion, achieving good accuracy. Using Grad-CAM heat maps, we can visualize that the AI model is focusing on the invasion areas.

Development Section in Charge

AI Development for Diagnostic Imaging Assistant Section

Principal Investigator

FUJIMA Noriyuki, Senior Lecturer, Department of Diagnostic and Interventional Radiology,
Hokkaido University Hospital

University-Affiliated Researchers

NAKAGAWA Junichi, PhD Student, Department of Diagnostic Imaging, Hokkaido University Graduate School of Medicine
HIRATA Kenji, Associate Professor, Department of Diagnostic Imaging, Hokkaido University Graduate School of Medicine
TANG Minghui, Specially Appointed Assistant Professor, Department of Diagnostic Imaging, Hokkaido University Graduate School of Medicine
KUDO Kohsuke, Professor, Department of Diagnostic Imaging, Hokkaido University Graduate School of Medicine