3D Endoscopic Lesion Diagnosis Using Point Cloud Networks: A Deep Learning Approach with PointNet

Inhalt:
This paper presents a novel deep learning framework that leverages PointNet for the classification of gastrointestinal lesions, including normal tissue, polyps, and esophagitis, by using 3D point cloud representations derived from endoscopic images. Unlike traditional 2D convolutional neural networks, our approach captures the complete 3D spatial geometry of lesions by reconstructing depth maps from 2D images and transforms them into point clouds. The proposed method comprises 3D preprocessing steps, including mesh cleaning, data augmentation, and classification using a robust PointNet architecture with T-Net transformation layers and dropout regularization. Experiments were conducted on a modified Kvasir v1 dataset, showing a classification accuracy of 92%, and high precision and recall across all categories. Comparative evaluations indicate superior performance, especially in capturing subtle structural differences. Confusion matrix analysis confirms the model's strong generalization and limited misclassification. This study highlights the potential of Point Cloud Networks in enhancing diagnostic accuracy, interpretability, and robustness in medical image analysis and processing. Future work includes integration with real-time endoscopic navigation systems and exploring advanced volumetric reconstruction methods to improve 3D fidelity.