Fig. 5

Typical cases of our DLR model that misled radiologists to make incorrect decisions. The top panel shows two PDAC lesions. All radiologists consider these two lesions to be PDAC lesions in the first reading. However, with access to the information from the DLR model, Reader-5 changed to the correct decision, considering them as CP lesions. In these two cases, the score of DLR model for PDAC is significantly higher than that of CP, and the area of the highlighted regions in the heatmaps are large and mostly distributed inside the tumor, which is consistent with other PDAC cases. Since Reader-5 is a junior radiologist, we believe that Reader-5’s mistakes may be due to lack of experience or carelessness. The bottom panel shows two CP lesions, which are inconsistent with the diagnosis of the radiologists in the first reading. However, with access to the information provided by the DLR model, all radiologists make the wrong decision. In these two cases, the misjudgment in the first case may be due to the large highlighted area of the generated heatmap, which is relatively rare in CP lesions, although most of the highlighted areas are still located at the boundary of the image. In the second case, the PDAC score with the DLR model is significantly higher than that of CP, which represents a case of AI misjudgment, thus misleading the radiologists. PDAC, pancreatic ductal adenocarcinoma; CP, chronic pancreatitis; ROI, region of interest; AI, artificial intelligence; DLR, deep learning radiomics