Improved Osteoporosis Prediction in Breast Cancer Patients Using a Novel Semi-Foundational Model

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Small patient cohorts are common in medical imaging research, hindering generalizable machine learning model development due to data sharing limitations and privacy concerns.
• Foundational models, pre-trained on large datasets, are emerging as a solution, particularly in natural language processing, but remain less developed for computer vision tasks.
• Protecting health information while enabling AI development necessitates innovative approaches like pretraining on native data feature spaces.

Purpose of the Study:

• To investigate the efficacy of pretraining established computer vision models (VGG-16, ResNet-50, DenseNet-121) on an unrelated chest CT dataset.
• To fine-tune these semi-foundational models for classifying bone mineral density (BMD) in breast cancer patients.
• To evaluate the performance improvement compared to traditional methods and ImageNet transfer learning.

Main Methods:

• Pretrained Visual Geometry Group (VGG)-16, Residual Network (ResNet)-50, and Dense Network (DenseNet)-121 models on 8500 chest CT scans.
• Fine-tuned the pre-trained models to classify bone mineral density (BMD) into mild, moderate, and severe categories using L1 vertebra CT scans from 199 breast cancer patients.
• Compared classification performance against ground truth Hounsfield Unit (HU) measurements and assessed statistical significance using ANOVA.

Main Results:

• Semi-foundational models demonstrated significantly improved ternary classification of BMD compared to Hounsfield Unit measurements.
• The ResNet-50 architecture, pre-trained on chest CTs, achieved the best performance.
• Holdout testing showed an Area Under the Curve (AUC) of 0.99 (p < 0.05) and an F1-score of 0.99 (p < 0.05).

Conclusions:

• Pretraining computer vision models on native CT data feature spaces enhances classification performance for disparate disease states, even with limited data.
• Semi-foundational models offer a promising strategy to overcome data scarcity and privacy issues in medical imaging AI.
• This approach has the potential for improved generalization in clinical applications with smaller disease-specific datasets.