Fig. 3: Visualization and interpretability of model predictions of high TIL infiltrated WSIs. | Communications Engineering

Fig. 3: Visualization and interpretability of model predictions of high TIL infiltrated WSIs.

From: Annotation-efficient deep learning for breast cancer whole-slide image classification using tumour infiltrating lymphocytes and slide-level labels

Fig. 3

Two example slide images with high TIL infiltration are presented in the left and right columns. Here we compare (a) the manual annotations of tumor (yellow line) and stroma (blue line) regions with the model predictions from (b) CLAM (c) CLAM (MoCo) and (d) ANSAC for four different threshold settings, \(t\) (where \(t=\left(i\right)\, 0,\left({ii}\right)\, 0.4,\left({iii}\right)\, 0.6{and}\left({iv}\right)\, 0.8\)). In the heatmaps, red regions have the highest attention, green regions have medium attention and blue regions have the lowest attention values (a color bar is provided at the bottom).

Back to article page