Optimization of Transfer Learning of Foundation Models for Hyperspectral Histologic Imaging.

Hyperspectral imaging (HSI) is a promising modality for digital pathology, but it is not yet widely adopted compared to traditional red-green-blue (RGB) histologic imaging. This study aims to develop techniques for transferring knowledge from histopathological foundation models trained on conventional RGB image datasets to models that can process data acquired by hyperspectral imaging. We used a dataset of 89 whole-slide hyperspectral histologic images from 54 patients to fine-tune three different foundation models. We also performed a hyperparameter search for each model and technique to identify general hyperparameter combinations well-suited for this task. Our results show that performing end-to-end fine-tuning of models generally outperforms other knowledge transfer paradigms, and that low learning rates and high weight decays tend to perform best for the transfer learning process. These findings partially contradict the common wisdom of first performing training only in the embedding layer, where gradients are concentrated. This study demonstrates a set of effective techniques for applying foundation models trained on RGB images to hyperspectral images for computational histopathology.