SP47. Artificial Intelligence Based Classification Of ICG Lymphography Patterns
Berk Barış Özmen, Sonia Kukreja Pandey, Graham S. Schwarz · Plastic & Reconstructive Surgery Global Open · 2024
Purpose: Effective imaging of the lymphatic system is paramount for comprehensive diagnosis of lymphedema. Indocyanine green (ICG) lymphography has become an important tool in diagnosis for lymphedema. With the widespread use of ICG lymphography to identify lymphatic flow abnormalities, there arises a need for objective, reproducible analysis methods. The purpose of this study is to investigate the accuracy of Artificial Intelligence (AI) Computer Vision Convolutional Neural Networks (CNN) algorithms for predicting ICG lymphography patterns. Methods: A dataset comprising 68 ICG lymphography images was compiled and labeled according to 5 recognized pattern types; linear, reticular, splash, stardust and diffuse. Dataset is partitioned into a training set (80%), validation set (10%) and test set (10%) all randomly. An AI CNN model was designed and trained on this dataset utilizing Python 3 and TensorFlow. Leveraging the power of transfer learning, we employed the MobileNetV2 architecture, pre-trained on the extensive ImageNet dataset. Custom dense layers were appended to this base model to adapt it to our specific classification task. To improve the model’s generalization capabilities, data augmentation techniques including rotation, width shift, height shift, shear transformations, zoom, and horizontal flipping were applied to the training data. Results: The AI model achieved an accuracy of 97.78%, with a loss of 0.0678 during 50 epochs of model training. Conclusion: Integration of AI computer vision models to ICG lymphography holds significant potential for enhancing the diagnostic accuracy and reproducibility of ICG lymphography. By refining the interpretation process of ICG lymphography with AI, we anticipate better pre-operative decision making for lymphedema, improved clinical management, and overall enhanced patient outcomes.