search

shartoo / luna16_multi_size_3dcnn

An implement of paper "Multi-level Contextual 3D CNNs for False Positive Reduction in Pulmonary Nodule Detection"
111 stars 45 forks source link

question about training #1

Closed dearkafka closed 6 years ago

dearkafka commented 6 years ago

It seems that as well as in paper, this implementation supposes that the nodule is in the center of the cropped cube. No shifting the center is provided. But in real life we really do not know if nodule is in center of the crop that we make from the scan, correct me if I'm wrong, but this is not directly useful for diagnosing.

shartoo commented 6 years ago

Yes.you're right.There is no absolute accurate coordinates,all are average values(coordinates are labeled by different radiation technologists). we'll usually do data enhancement by shifting on center for deep learning task.

dearkafka commented 6 years ago

Thank you very much! Do you know any baseline, published or your own, for this kind of task on this LIDC dataset but with shifting center? I mean any usual metrics, i.e. precision, recall? I do this kind of task, mainly segmentation, achieve high dice score of 0.8-0.9, however, unstable on different batches, but sensitivity looks rather low, like 0.7

shartoo commented 6 years ago

Sorry,i'm not focus on lung cancer detecttion and just know few of this kind of paper.Here is the link of papers using LIDC

dearkafka commented 6 years ago

Thank you for your help, forgot about this list!