Yasmina Al Khalil

Fine-tuning pre-trained models has become a cornerstone of modern machine learning, allowing practitioners to achieve high performance with limited labeled data. In surgical video analysis, where expert annotations are especially time-consuming and costly, identifying the most suitable pre-trained model for a downstream task is both critical and challenging. Source-independent transferability estimation (SITE) offers a solution by predicting how well a model will fine-tune on target data using only its embeddings or outputs, without requiring full retraining. In this work, we formalize SITE for surgical phase recognition and provide the first comprehensive benchmark of three representative metrics, LogME, H-Score, and TransRate, on two diverse datasets (RAMIE and AutoLaparo). Our results show that LogME, particularly when aggregated by the minimum per-subset score, aligns most closely with fine-tuning accuracy; H-Score yields only weak predictive power; and TransRate often inverses true model rankings. Ablation studies show that when candidate models have similar performances, transferability estimates lose discriminative power, emphasizing the importance of maintaining model diversity or using additional validation. We conclude with practical guidelines for model selection and outline future directions toward domain-specific metrics, theoretical foundations, and interactive benchmarking tools.

We propose a method for synthesizing cardiac magnetic resonance (MR) images with plausible heart pathologies and realistic appearances for the purpose of generating labeled data for the application of supervised deep-learning (DL) training. The image synthesis consists of label deformation and label-to-image translation tasks. The former is achieved via latent space interpolation in a VAE model, while the latter is accomplished via a label-conditional GAN model. We devise three approaches for label manipulation in the latent space of the trained VAE model; i) intra-subject synthesis aiming to interpolate the intermediate slices of a subject to increase the through-plane resolution, ii) inter-subject synthesis aiming to interpolate the geometry and appearance of intermediate images between two dissimilar subjects acquired with different scanner vendors, and iii) pathology synthesis aiming to synthesize a series of pseudo-pathological synthetic subjects with characteristics of a desired heart disease. Furthermore, we propose to model the relationship between 2D slices in the latent space of the VAE prior to reconstruction for generating 3D-consistent subjects from stacking up 2D slice-by-slice generations. We demonstrate that such an approach could provide a solution to diversify and enrich an available database of cardiac MR images and to pave the way for the development of generalizable DL-based image analysis algorithms. We quantitatively evaluate the quality of the synthesized data in an augmentation scenario to achieve generalization and robustness to multi-vendor and multi-disease data for image segmentation. Our code is available at https://github.com/sinaamirrajab/CardiacPathologySynthesis