Figure 3 Confusion matrices for the binary (right side) and
multiclass (left side) classification between the benign and malignant
lymphoid tissues. The x-axis shows the true label. The y-axis
corresponds to the predicted label.
As shown in Figure 3, the accuracy is above 95% for every pair
of malignant lymphoid tissue and still above 90% for the distinction
between benign and malignant lymphoid tissue. The multiclass
classification between the lymphoma types (FL and DLBCL) and
rLNFC and rLNMZ leads to similarly
high accuracy for the binary classification. This multiclass
classification’s accuracy is still above 94 % for all neoplastic and
reactive lymphoid tissue types.
The neural network, once trained, performs a forward execution of a
complex function depicted by the NN, representing the analysis of an IR
image in less than 25 ms. NNs with three output classes have been
trained to visualize the capabilities of the deep learning approach.
These three classes serve as a colour code for an RGB image, whereby the
NN classifies the spectrum of each pixel. This means that it was decided
to which class the corresponding subsection belongs for each pixel
individually. This leads to the satisfactory resolution images shown inFigure 4 A-C .