Deep-learning framework for analysis of longitudinal MRI studies

Longitudinal imaging studies, characterized by scans of an individual at multiple time points, are essential for diagnosis, treatment planning, and tracking the response to therapy. However, the complexity of reading and the volume of follow-up studies, along with other factors, make the analysis of longitudinal studies a time-consuming and challenging task for radiologists. These challenges highlight the need for automated data analysis methods that can aid radiologists in interpreting longitudinal images. Recent advancements in deep learning hold promise for the development of such methods. We propose a novel deep learning framework for the automated analysis of longitudinal data to detect significant changes, specifically targeting tumor growth assessment. Our approach employs a Siamese network architecture with a shared convolutional neural network (CNN) backbone, facilitating the direct comparison of features from pairs of input volumetric images and predicting the degree of change as the output. We validated our framework on both synthetic and real datasets. In evaluations on the synthetic dataset, the model achieved a Pearson correlation coefficient, Area Under the Curve (AUC), and <i>F</i>₁ score of 0.98, 0.98, and 0.97, respectively. In our evaluations on the real dataset, the model demonstrated an AUC of 0.83 and an <i>F</i>₁ score of 0.86. Our method effectively captures significant changes in tumor size between pairs of input images, operates efficiently with limited labeled data, and offers interpretable visualizations via gradient-based class activation maps. While challenges remain in aligning the performance between synthetic and real data, our findings underscore the potential of deep learning to significantly aid the analysis of longitudinal studies within the radiology workflow.