Fig. 5: Deep learning pipeline. | npj Precision Oncology

Fig. 5: Deep learning pipeline.

From: An interpretable deep learning model for predicting endometrial cancer molecular subtypes from H&E-stained slides

Fig. 5: Deep learning pipeline.The alternative text for this image may have been generated using AI.

Whole slide image (WSI) was cut into 1024 × 1024-pixel tiles and the non-tissue tiles were excluded using the histogram method. The tissue tiles were passed to a DeepLabv3 model and classified into tumor regions(red) and normal regions(blue). The tumor tiles were color normalized using the Vahadane method and converted into mini-batches. Our proposed prediction network, based on EfficientNetV2, was trained to capture global shapes and local texture features from the tiles and finally output tile-level prediction probabilities for the four molecular subtypes. Hover-Net, a nuclear segmentation and classification deep learning model, was used to analysis the multiple features in ROI on 364 WSIs from Fudan Cohort. A WSI-level feature set was established by aggregating the tile-level features, including basic information, morphological features, texture features and spatial distribution. MSI-H microsatellite instability–high. NSMP no specific molecular profile. p53abn abnormal cellular tumor antigen p53 expression.

Back to article page