Explainable Deep Learning for Automated Classification of Skeletal and Tumor Pathologies in 3D Volumetric Medical Imaging.
Authors
Affiliations (7)
Affiliations (7)
- Department of Computer Engineering and Networks, College of Computer and Information Sciences, Jouf University, Sakaka, 72388, Saudi Arabia.
- Department of Electrical and Computer Engineering, College of Engineering, Anderson University, Anderson, 29621, SC, USA.
- School of Medicine Greenville, University of South Carolina, Greenville, 29605, SC, USA.
- Department of Computer Science, College of Computer Engineering and Sciences, Prince Sattam Bin Abdulaziz University, Alkharj, 11942, Saudi Arabia.
- School of Computer Science and Engineering, Lovely Professional University, Phagwara, 144411, Punjab, India.
- College of Computing and Information Technologies, National University, Manila, Philippines.
- School of Management and Information Technology, De La Salle-College of Saint Benilde, Manila, 1004, Philippines. [email protected].
Abstract
Accurate identification and classification of 3D mesh data in medical imaging are crucial for various clinical and research applications, including surgical planning, anatomical analysis, and disease diagnosis. In recent years, the application of deep learning techniques to medical imaging has gained significant traction, particularly in the analysis of 3D mesh data. This study presents a comprehensive explainable AI-based deep learning framework for classifying tumorous and skeletal structures in 3D medical data. The models, including CNN, ResNet-18, VGG-16, and EfficientNet-B0, are trained and evaluated on MedShapeNet, a large-scale dataset comprising over 100,000 3D medical shapes spanning bones, organs, vessels, muscles, and surgical instruments. A robust preprocessing pipeline is developed to transform raw Stereolithography (STL) neuro-anatomical structures into normalised feature vectors, ensuring consistent representation across varying mesh complexities. A multifaceted model evaluation approach is employed, combining quantitative metrics with advanced visualisation techniques to gain insight into the model's decision-making processes. Experimental results show that the VGG-16 model achieved the highest test accuracy of 99.15% and an F1 score of 99.16%, closely followed by ResNet-18 with an accuracy of 98.87%. These findings highlight the effectiveness of VGG in improving deep learning performance for medical mesh classification tasks.