zhixuhao / unet

unet for image segmentation
MIT License
4.59k stars 2k forks source link

failed to work when using my own dataset #154

Open lu-jincheng opened 5 years ago

lu-jincheng commented 5 years ago

I can generate good result when training and testing with the provided image data, but when turning to my own dataset, which were chest X-ray images with lung masks, the test result images were basicly all black, sometimes with a white line near the edge. The metric "accuracy" was above 0.9 during training but the output images were bad, that is what I'm confused aboout. I tried to replace the metrics with dice, but the result were still all black.

synioe commented 5 years ago

Maybe there are some mistakes in you own dataset.By the way,I'm working on training my own dataset.How did you make labels with you own dataset?

happy20200 commented 4 years ago

在使用提供的图像数据进行训练和测试时,我可以产生良好的结果,但是当转到我自己的数据集时,即带有肺罩的胸部X射线图像,测试结果图像基本上都是黑色的,有时在边缘。训练期间的指标“准确性”高于0.9,但输出图像很差,这就是我感到困惑的地方。我尝试用骰子替换指标,但结果仍然是黑色的。 May I ask if you solved this problem, I was testing directly with my own data but it didn't work well. image