A Lightweight Machine Learning Model for High Precision Gastrointestinal Stromal Tumors Identification

Gastrointestinal stromal tumors (GISTs), which usually develop with a significant malignant potential, are a serious challenge in stromal health. With Endoscopic ultrasound (EUS), GISTs can appear similar to other tumors. This study introduces a lightweight convolutional neural network model optimized for the classification of GISTs and leiomyomas using EUS images only. Models are constructed based on a dataset that comprises 13277 augmented grayscale images derived from 703 patients, ensuring a balanced representation between GIST and leiomyoma cases. The optimized model architecture includes seven convolutional units followed by fully connected layers. After being trained and evaluated with a 5-fold cross-validation, the optimized model achieves an average validation accuracy of 96.2%. The model achieved a sensitivity, specificity, positive predictive value, and negative predictive value of 97.7%, 94.7%, 94.6%, and 97.7%, respectively, and significantly outperformed endoscopists' assessments. The study highlights the model's robustness and consistency. Our results suggest that instead of using developed deep models with fine-tuning, lightweight models with their simpler designs may grasp the essence and drop speckle noise. A lightweight model as a hypothesis with fewer model parameters is preferable to a deeper model with 10 times the model parameters according to Occam's razor statement.