Fig. 1: The framework of MILTS and its performance on clinically PDL1-relevant tumors. | Nature Communications

Fig. 1: The framework of MILTS and its performance on clinically PDL1-relevant tumors.

From: Teacher-student collaborated multiple instance learning for pan-cancer PDL1 expression prediction from histopathology slides

Fig. 1

a The training and inference workflow of MILTS includes three steps. First, the data of patient cohorts are divided into training set, validation set and test set, followed by patching and random augmentations. Then, obtained tiles are utilized to train the patch-level teacher-student collaborated network in a MIL manner. At last, the trained patch-level teacher model (or student model) works as the extractor of both statistical features and deep features. The deep features of patches in the same slide are further fused into a slide token and combined with the statistical summary of patch-level features to train an MLP classifier which gives the patient-level diagnosis. MIL multiple instance learning, S student, T teacher, C concatenation, MLP multi-layer perceptron. b Quantities of slide images of different tumors. c Plot illustrating the model’s performance on FFPE slides and fresh-frozen slides for the aforementioned tumors, separately. Source data are provided as a Source Data file.

Back to article page