Prediction of cervix cancer stage and grade from diffusion weighted imaging using EfficientNet

Area of Science:

• Radiology and Medical Imaging
• Oncology
• Artificial Intelligence in Medicine

Background:

• Cervix cancer (CC) grading and staging are crucial for treatment planning.
• Current methods may be invasive or lack precision.
• Diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps offer potential for noninvasive assessment.

Purpose of the Study:

• To develop and validate a noninvasive method for predicting CC grade and stage using a single image.
• To leverage deep convolutional neural networks (DCNN) for image analysis of ADC maps.
• To assess the performance of DCNN models against traditional radiomic analysis.

Main Methods:

• Retrospective collection of ADC map data from 85 CC patients.
• Extraction of gross tumor volume (GTV) slices from pre-treatment DWI.
• Application of EfficientNetB0 and EfficientNetB3 models for binary and four-class classification tasks.
• Utilized Synthetic Minority Oversampling Technique (SMOTE) for balanced training sets and five-fold cross-validation.

Main Results:

• EfficientNetB3 achieved an AUC of 0.924 for grade prediction.
• EfficientNetB0 achieved an AUC of 0.931 for stage prediction.
• DCNN models significantly outperformed radiomic analysis (AUCs of 0.67-0.66) and showed superior or comparable results to ResNet50 and Xception.

Conclusions:

• Predicting CC grade and stage from ADC maps using DCNN is feasible.
• EfficientNet-based approaches demonstrate high accuracy in noninvasive CC assessment.
• This method offers a promising tool for improved CC diagnosis and treatment stratification.