Deep Convolutional Neural Network for Accurate Classification of Myofibroblastic Lesions on Patch-Based Images

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Digital pathology

Background:

• Myofibroblastic lesions require accurate classification into benign and malignant categories for appropriate patient management.
• Distinguishing between benign and malignant myofibroblastic tumors can be challenging due to histological similarities.

Purpose of the Study:

• To implement and evaluate a Deep Convolutional Neural Network (CNN) for classifying myofibroblastic lesions.
• To assess the performance of a ResNet50 model in differentiating benign from malignant myofibroblastic tumors using patch-based image analysis.

Main Methods:

• A Residual Neural Network (ResNet50) model was fine-tuned using ImageNet pre-trained weights.
• Whole-slide images (WSIs) were segmented into 224x224 pixel patches and divided into training, validation, and test sets.
• The CNN was trained for 75 epochs with the Adam optimizer and a learning rate of 0.00001.

Main Results:

• The ResNet50 model achieved high performance metrics: 98.97% accuracy, 99.91% precision, 97.98% sensitivity, 99.91% specificity, 98.94% F1 score, and 0.99 AUC.
• The model demonstrated robust generalization capabilities on unseen data.

Conclusions:

• The developed ResNet50 model shows nearly flawless performance in distinguishing benign from malignant myofibroblastic tumors.
• The AI model's effectiveness may stem from its ability to identify subtle discriminative features and leverage comprehensive data preprocessing.