Effect of Dataset Size and Medical Image Modality on Convolutional Neural Network Model Performance for Automated...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Accurate medical image segmentation is vital for diagnosing and treating diseases like renal tumors.
• The optimal size of training datasets for deep learning models in medical imaging remains a challenge.
• Previous studies have not established a clear method for predicting dataset size requirements for peak model performance.

Purpose of the Study:

• To investigate the efficacy of an exponential-plateau model for determining the optimal training dataset size for medical image segmentation.
• To establish the relationship between dataset size and model generalizability performance across different imaging modalities and target structures.
• To provide a predictive tool for researchers to ascertain when additional training data will not significantly improve model performance.

Main Methods:

• Retrospective collection of CT and MR images of patients with renal tumors.
• Assembly of modality-based datasets with varying image counts (50-300) for model training and validation (80-20 split).
• Evaluation against a held-out test set (50 images) and analysis using an exponential-plateau model to identify performance plateaus.

Main Results:

• The exponential-plateau model successfully predicted performance plateaus for segmenting non-neoplastic kidney regions and tumor regions on CT and MR images.
• Specific numbers of training-validation images required to reach performance plateaus were identified for different segmentation tasks (e.g., 54 for non-neoplastic kidney on CT, 389 for tumor on MR).
• Experiments with the KiTS21 dataset and different model architectures (nn-UNet 2D/3D) confirmed that modality, target structure, and architecture influence required dataset size.

Conclusions:

• The developed exponential-plateau modeling approach effectively predicts the dataset size needed to achieve maximum performance in medical image segmentation.
• The required number of training images varies significantly based on imaging modality, the specific anatomical structure being segmented, and the chosen model architecture.
• This methodology offers a valuable tool for researchers to optimize data acquisition and model training, preventing unnecessary data collection and improving efficiency.