Ovarian tumor diagnosis using deep convolutional neural networks and a denoising convolutional autoencoder

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate ovarian tumor classification is crucial for effective treatment planning.
• Current diagnostic methods may benefit from enhanced computational tools for improved accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate a deep learning model, CNN-CAE, for classifying ovarian tumors using ultrasound images.
• To assess the model's ability to remove image artifacts and accurately discriminate between normal ovaries, benign tumors, and malignant tumors.

Main Methods:

• A CNN-CAE model was developed using 1613 pre-processed and augmented ovarian ultrasound images.
• Fivefold cross-validation was employed to evaluate model performance metrics including accuracy, sensitivity, specificity, and AUC.
• Gradient-weighted class activation mapping (Grad-CAM) was utilized for qualitative verification of model predictions.

Main Results:

• The CNN-CAE model achieved 97.2% accuracy and 0.9936 AUC in classifying normal versus ovarian tumors using DenseNet121.
• For distinguishing malignant ovarian tumors, the model demonstrated 90.12% accuracy and 0.9406 AUC with DenseNet161.
• Grad-CAM analysis confirmed the model's reliance on relevant texture and morphology features for classification.

Conclusions:

• The CNN-CAE model is a feasible and robust tool for classifying ovarian tumors from ultrasound images, effectively handling image artifacts.
• The model exhibits significant clinical application value, potentially improving diagnostic accuracy and patient management in oncology.