Explored the application of physics-based deep learning for fast inference of molecular composition of brain tissue from spectroscopy- based imaging modality such as hyperspectral imaging and near- infrared spectroscopy. This method is based on the paper "Shallow learning enables real-time inference of molecular composition changes from broadband-near-infrared spectroscopy of brain tissue" by IVAN EZHOV, LUCA GIANNONI, IVAN ILIASH, FELIX HSIEH, CHARLY CAREDDA, FRED LANGE, ILIAS TACHTSIDIS, AND DANIEL RUECKERT.
-
optimisation_diff_batch.ipynb - Demonstrate the method on the piglet dataset. Tissue concentration of "HbO2", "Hbb", "oxyCCO", "redCCO" are derived from the spectra using the cvpx library as ground truth for the neural network.
-
inference_time.ipynb - Various neural netork architecure are training, their performance is compared in this notebook to rank them.
-
optimisation_diff_batch_simulation.ipynb - Demonstrate our method on the mouse dataset that is simulated using monte carlo method. We employed a small neural network with 2 layers and 256 hidden units. The plotting shows the activation map of each of the 4 tissues concentration predicted, one predicted using the neural network and another the groud truth derived from the optimization method.
-
preprocessing.py - To generate the molecular coefficient of the 4 tissue molecules by interpolating the input spectrato generate the system matrix for the optimization method.
-
train_mouse.py - Training script for the mouse dataset demonstrated in optimisation_diff_batch_simulation.ipynb.
-
train_piglet.py - Training script for the piglet dataset demonstrated in optimisation_diff_batch.ipynb.