A deep learning-based cancer survival time classifier for small datasets.

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep Learning (DL) shows promise for cancer survival time prediction.
• Limited annotated medical imaging data hinders DL model performance in clinical settings.
• Overfitting is a significant challenge when training DL models on small datasets.

Purpose of the Study:

• To develop a customized neural network model for effective DL training in small sample spaces.
• To improve the reliability of prognostic radiomic feature selection for cancer survival prediction.
• To enhance the clinical utility of DL models for lung tumor evaluation and patient care.

Main Methods:

• A neural network model was customized for small sample sizes to prevent data overfitting.
• Prognostic radiomic features were selected iteratively using the average of multiple dropouts.
• The proposed classifier was compared against the erasing feature selection method and other classifiers on small sample data.

Main Results:

• The proposed method demonstrated improved network learning capability and reliable feature selection.
• Statistically validated results showed efficient and improved classification of cancer survival time into three intervals (≤6 months, 6 months–2 years, >2 years).
• The model achieved better training performance over a small database compared to existing methods.

Conclusions:

• The customized neural network model effectively addresses the challenge of limited data in DL-based cancer survival prediction.
• The proposed feature selection method enhances network learning and classification accuracy.
• This approach has the potential to aid healthcare professionals in lung tumor evaluation for timely treatment and patient management.