Fig. 8: Aggregation operations in MIL and in whole-slide training method. | Nature Communications

Fig. 8: Aggregation operations in MIL and in whole-slide training method.

From: An annotation-free whole-slide training approach to pathological classification of lung cancer types using deep learning

Fig. 8: Aggregation operations in MIL and in whole-slide training method.The alternative text for this image may have been generated using AI.

a In multiple-instance learning (MIL), a “max” operation is performed to select the most representative patches by their mapped predicted scores. b In the whole-slide training method, the front part of the model encodes the entire image into an embedding feature map, where each vector along the channel axis corresponds to a receptive field. The following GMP (or GAP) layer embeds 2048 max (or average) operations to reduce the spatial dimension. Then, a dense layer performs a linear transformation followed by an activation function on the reduced 2048-length vector for a slide-level prediction.

Back to article page