Fig. 7.

Performance for the training process of detecting nodules. (a): The dichotomy loss embodies the ability to judge whether nodules existed in the mass of anchor frames generated by the detection model. If the overlapping area of the anchor frame and the real nodule box was larger than a certain threshold, a nodule in the anchor frame shall be considered. (b): Position regression loss is harnessed to assess the accuracy of the detection frame position, aiming to make the model-based detection frame close to the real nodule frame, with the same premise as nodule classification loss. (c): The classification loss reflects the ability to determine the nodule category, such as 0–3 mm nodules and 3–6 mm nodules. The premise is that at least one nodule existed in the anchor box and failed otherwise. (d): The accuracy of nodule detection refers to the dichotomic ability of accurately distinguishing nodules and backgrounds. (e): Nodule detection recall rate indicates the number of nodules found in the model compared to the total number, i.e., the recall rate of the model.