Clarifying how foundation model encoders change during fine-tuning is important for transparency and trustworthiness in their medical imaging applications. Using prostate cancer patient imaging data, we train a two-layer MLP to learn the relationship between radiomic descriptors and encoder embeddings prior to fine-tuning. This model captures non-linear associations through its first layer, while the final linear layer offers an interpretable mapping from radiomic attributes to (transformed) latent features. To quantify the effect of fine-tuning, the first layer is fixed, and only the linear layer is re-estimated using the embeddings from the fine-tuned encoder. Comparing the pre- and post-fine-tuning linear weights yields a direct quantitative measure of how the encoder’s emphasis on specific radiomic characteristics shifts during fine-tuning.
We validate the approach using a prostate MRI foundation model (ProFound) and multiple downstream tasks.
