Ranking attention multiple instance learning for lymph node metastasis prediction on multicenter cervical cancer MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Current diagnosis of lymph node metastasis (LNM) relies on invasive histopathological examination after surgery.
• There is a critical need for non-invasive, preoperative methods to predict LNM, especially in cervical cancer.
• Accurate LNM prediction is crucial for effective treatment planning and patient outcomes.

Purpose of the Study:

• To develop and validate a novel non-invasive method for preoperative prediction of lymph node metastasis (LNM) in cervical cancer patients.
• To integrate deep learning with attention mechanisms for enhanced diagnostic accuracy using T2-weighted MRI.
• To provide visual interpretability for the model's predictions, identifying informative regions associated with LNM.

Main Methods:

• A Ranking Attention Multiple Instance Learning (RA-MIL) model was developed, combining Convolutional Neural Networks (CNNs) with ranking attention pooling.
• The RA-MIL model processes 2D MRI slices to extract imaging features and generates patient-level representations for LNM classification.
• The study included 300 female patients with cervical cancer, utilizing data from both a hospital and an open-source dataset.

Main Results:

• The RA-MIL model achieved a high area under the receiver operating characteristic curve (AUC) of 0.809 on the internal test set and 0.833 on the public dataset.
• The model demonstrated significant improvements in LNM status prediction compared to other state-of-the-art deep learning models.
• Visualizations highlighted informative MRI slices and regions, enhancing the interpretability of the model's predictions.

Conclusions:

• The developed RA-MIL model shows potential as a non-invasive auxiliary tool for preoperative LNM prediction in cervical cancer.
• The model offers visual interpretability, aiding clinicians in understanding the basis of the diagnostic predictions.
• This approach could improve diagnostic accuracy and guide treatment decisions, reducing the need for invasive procedures.