Fig. 2: Receiver operating characteristic (ROC) curves of the metastasis identification ability of the models under the main test set of 1156 LN images. | Nature Communications

Fig. 2: Receiver operating characteristic (ROC) curves of the metastasis identification ability of the models under the main test set of 1156 LN images.

From: Deep neural network trained on gigapixel images improves lymph node metastasis detection in clinical settings

Fig. 2

The rows present a comparison of (a) the performance of our model with those of three pathologists before and after receiving AI assistance on the 48 equivocal cases; (b) weakly supervised methods; (c) performance obtained under various magnification levels (20×: 0.46 µm/pixel, 10×: 0.92 µm/pixel, 5×: 1.84 µm/pixel); and (d) performance obtained under different label types and amounts of training data. We evaluated each method with the main test set and its subsets to retrieve the ROC curves. The first column presents the ROC curves differentiating between the 1156 LNs in the main test set. Among the 1156 LNs, those marked as micrometastases and ITCs, as well as all the negative LNs, were sampled to evaluate the performance of the model in identifying micrometastases and ITCs, as presented in the second and third columns. The fourth column displays the slide-level performance.

Back to article page