“…A multi-layer perceptron was then trained on these supertiles to predict gene expression data. During training, only the top k [ 1 , 2 , 5 , 10 , 20 , 50 ] , 100 tiles in terms of predicted gene expression are selected and a weighted mean computed of their predictions, giving more weight to higher valued predictions. Finally, the model is fine-tuned for specific organs using the 8000 normal tiles, again sampling the top k [ 10 , 20 , 50 ], 100, 200, 500, 1000, 2000, 5000 during training.…”