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

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Gastrointestinal stromal tumors (GISTs) pose diagnostic challenges due to their potential malignancy and resemblance to other tumors in endoscopic ultrasound (EUS) imaging.
• Accurate differentiation between GISTs and leiomyomas is crucial for appropriate patient management.

Purpose of the Study:

• To develop and evaluate a lightweight convolutional neural network (CNN) model for classifying GISTs and leiomyomas using only EUS images.
• To assess the performance of the lightweight CNN model against human expert assessments.

Main Methods:

• A dataset of 13,277 augmented grayscale EUS images from 703 patients was used, ensuring a balanced representation of GIST and leiomyoma cases.
• A lightweight CNN architecture with seven convolutional units and fully connected layers was designed and optimized.
• The model was trained and evaluated using 5-fold cross-validation.

Main Results:

• The optimized lightweight CNN model achieved an average validation accuracy of 96.2%.
• The model demonstrated high performance metrics: 97.7% sensitivity, 94.7% specificity, 94.6% positive predictive value, and 97.7% negative predictive value.
• The AI model significantly outperformed the diagnostic accuracy of endoscopists.

Conclusions:

• Lightweight CNN models, with their simpler designs, can effectively capture essential image features and reduce noise for accurate GIST classification.
• The proposed lightweight model offers a robust and consistent alternative to more complex deep learning models, aligning with Occam's razor principle.
• This AI-driven approach shows significant potential for improving the diagnostic accuracy of GISTs and leiomyomas in EUS imaging.