Democratizing Artificial Intelligence in Healthcare: A Study of Model Development Across Two Institutions Incorporating Transfer Learning

Vikash Gupta, Holger Roth, Varun Buch, Marcio Aloisio Bezerra Cavalcanti Rockenbach, Richard D. White, Dong Yang, Olga Laur, Brian B. Ghoshhajra, Ittai Dayan, Daguang Xu, Mona G. Flores, Barbaros Selnur ErdalarXiv 2020 · OpenReview

Introduction: Overcoming the Radiology Bottleneck

The integration of deep learning into clinical workflows is often hindered by the "Radiology Bottleneck." Unlike general computer vision tasks that benefit from millions of public images, medical AI requires expert radiologists to spend hundreds of hours manually segmenting complex 3D anatomy. This high barrier to entry prevents many institutions from developing bespoke models. This study investigates whether transfer learning (TL)—fine-tuning a model pre-trained at one institution for use at another—can democratize AI development by reducing the data and labor required for high-performance results.

Methodology

The researchers focused on segmenting the Left Ventricular Myocardium (LVM) from 4D Cardiac Computed Tomography Angiography (CCTA). These scans are particularly complex as they involve three spatial dimensions plus the temporal dimension of the beating heart.

The study compared two primary approaches at a target institution ("Site 2"):

  1. Random Initialization: Training a 4D CNN from scratch using only local data.
  2. Transfer Learning (TL): Utilizing a model pre-trained at "Site 1" and fine-tuning it with a limited local dataset from Site 2.

To simulate realistic clinical constraints, the fine-tuning used "sparse" labeling (annotating only two phases of the cardiac cycle), whereas the pre-training at Site 1 used "dense" labeling (annotating 10 phases).

Qualitative and Quantitative Results

The transfer learning approach demonstrated a significant advantage in both efficiency and accuracy. Quantitatively, the TL model achieved a mean Dice coefficient of 0.8925, whereas the model trained from scratch reached only 0.6359, even when provided with three times more local training data.

Notably, the TL model reached a performance plateau with as few as 10 local cases. This suggests that once a model learns the fundamental features of cardiac CT imaging, it requires very little site-specific data to adapt to new scanners and protocols.

Comparison of segmentation results between random initialization (top) and transfer learning (bottom) across axial, coronal, and sagittal views. Note the reduction in "holes" and artifacts in the TL model.

Visually, the TL model produced much cleaner segmentation masks. As shown in the comparative views, the model trained from scratch often suffered from "holes" and gaps in the myocardium, while the TL-enhanced model produced smooth, physiologically plausible contours.

Clinical Validation via 3D Parameterization

To ensure the AI's output was clinically relevant, the team moved beyond simple pixel-overlap metrics. They employed a 3D parameterization technique based on heat transfer equations to calculate LVM wall thickness.

3D parameterization of the left ventricular myocardium showing streamlines and surface thickness maps used for clinical validation.

By comparing the AI-generated wall thickness to ground-truth clinical measurements, the researchers found a high Pearson’s Correlation (up to 0.92). This confirms that transfer learning doesn't just produce visually appealing masks; it generates data that accurately reflects the physiological reality of the patient's heart.

Conclusion

Transfer learning significantly lowers the threshold for implementing high-quality AI in radiology. By achieving superior results with just 10 local cases compared to 30 cases from scratch, this methodology allows institutions with limited resources to develop sophisticated diagnostic tools. While institutional variability and legal hurdles remain, the technical path toward democratizing medical AI is increasingly clear.

Summary updated

Abstract

The training of deep learning models typically requires extensive data, which are not readily available as large well-curated medical-image datasets for development of artificial intelligence (AI) models applied in Radiology. Recognizing the potential for transfer learning (TL) to allow a fully trained model from one institution to be fine-tuned by another institution using a much small local dataset, this report describes the challenges, methodology, and benefits of TL within the context of developing an AI model for a basic use-case, segmentation of Left Ventricular Myocardium (LVM) on images from 4-dimensional coronary computed tomography angiography. Ultimately, our results from comparisons of LVM segmentation predicted by a model locally trained using random initialization, versus one training-enhanced by TL, showed that a use-case model initiated by TL can be developed with sparse labels with acceptable performance. This process reduces the time required to build a new model in the clinical environment at a different institution.
BibTeX
@misc{gupta2020democratizing,
  author = {Vikash Gupta and Holger Roth and Varun Buch and Marcio Aloisio Bezerra Cavalcanti Rockenbach and Richard D. White and Dong Yang and Olga Laur and Brian B. Ghoshhajra and Ittai Dayan and Daguang Xu and Mona G. Flores and Barbaros Selnur Erdal},
  title = {{Democratizing Artificial Intelligence in Healthcare: A Study of Model Development Across Two Institutions Incorporating Transfer Learning}},
  year = {2020},
  eprint = {2009.12437},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2009.12437},
}
Figure