Figure 3
From: Perineural invasion detection in pancreatic ductal adenocarcinoma using artificial intelligence
Testing and Verification set graphs after elimination of labels less than 10,000 pixels (The algorithm was run on 168 independently labeled images the algorithm was not trained on): (a) intersection over Union curve: plots the percentage of overlap between the algorithm’s and the pathologist’s cancer and nerve markings against the confidence threshold. (b) Detection rate curve: plots the percentage of cancer and nerve that the algorithm detected against the confidence threshold. (c) False Alarm curve: plots the percentage of cancer and nerve that the algorithm falsely identified against the confidence threshold. Steadily decreasing IOU and detection rates are seen with increasing confidence threshold. Elimination of findings less than 10,000 pixels resulted in a flip where the nerves now have lower false alarm rates than tumor at confidence intervals greater than 18%.
