Optimizing malignancy prediction: A comparative analysis of transfer learning techniques on EBUS images

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Diagnostics
• Thoracic Medicine

Background:

• Improving diagnostic accuracy in Endobronchial Ultrasound (EBUS) image analysis using machine learning remains a significant challenge.
• Accurate prediction of lymph node malignancy is crucial for effective patient management in thoracic oncology.

Purpose of the Study:

• To identify the most effective transfer learning model for predicting lymph node malignancy from EBUS images.
• To evaluate the performance of various pre-trained Convolutional Neural Network (CNN) models for this diagnostic task.

Main Methods:

• Retrospective analysis of EBUS images collected between 2020-2023.
• Evaluation of eight pre-trained CNN models: VGG, ResNet, InceptionNet, Xception, MobileNet, DenseNet, NasNet, and EfficientNet.
• Assessment of model performance using metrics such as Area Under the Curve (AUC) and accuracy, with a focus on identifying overfitting.

Main Results:

• VGG19, EfficientNetV2L, and DenseNet201 demonstrated the highest performance in malignancy prediction, achieving AUCs of 0.96, 0.96, and 0.95, respectively.
• These top-performing models showed consistent training and testing accuracy, indicating successful generalization without overfitting.
• In contrast, ResNet152V2, Xception, and NasNet exhibited lower performance (AUCs 0.88-0.84) and signs of overfitting, while MobileNetV2 (AUC 0.50) failed to discriminate between benign and malignant cases.

Conclusions:

• Transfer learning applied to EBUS image analysis holds substantial promise for enhancing diagnostic accuracy in thoracic medicine.
• The study highlights specific CNN models that are highly effective for predicting lymph node malignancy, particularly in the context of lung cancer diagnosis.
• Optimized machine learning models can significantly aid clinicians in making more accurate and timely diagnoses.