Development and validation of a meta-learning-based multi-modal deep learning algorithm for detection of peritoneal...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Existing medical imaging struggles with early detection of small peritoneal metastases (PM), with accuracies as low as 29% for lesions under 0.5 cm.
• Early detection of PM is critical for timely intervention and improved patient outcomes.

Purpose of the Study:

• To develop a deep convolution neural network classifier utilizing meta-learning for enhanced prediction of peritoneal metastasis.
• To improve the accuracy and robustness of PM detection, particularly for smaller lesions.

Main Methods:

• A multi-modal deep Convolutional Neural Network (CNN) was trained using meta-learning techniques.
• Enhanced CT scans from plain, portal venous, and arterial phases were combined as multi-modal inputs.
• The model's performance was evaluated on a dedicated test dataset and compared against established PM prediction algorithms.

Main Results:

• The meta-learning based multi-modal PM predictor achieved an accuracy of 87.5% with an Area Under Curve (AUC) of 0.877.
• The classifier demonstrated high specificity (95.2%) and sensitivity (73.4%) on the testing datasets.
• Performance surpassed logistic regression (AUC: 0.795), ResNet3D (AUC: 0.827), and MADDG (AUC: 0.834).

Conclusions:

• A novel meta-learning training strategy was proposed, enhancing model robustness for predicting unseen peritoneal metastasis samples.
• The meta-learning based multi-modal classifier achieved superior results in synchronous PM prediction compared to existing algorithms.
• The study highlights the model's improved generalization ability, even with limited data, for effective PM detection.