I am using Mask-RCNN model with ResNet50 backbone for nodule detection in ultrasound images. My dataset consists of 500 US images. Maximum object detection accuracy for training set is approximately 54% (using data augmentation and hyper-parameter tuning). The accuracy of object detection on my test set is even lower. Are there any suggestions for improving object detection accuracy?
Many thanks in advance,
you can try different methods like yolo also you can use other loss functions to see how it works,watch this : https://lilianweng.github.io/lil-log/2018/12/27/object-detection-part-4.html
your dataset has only 500 sample which is very low,instead of resnet50 maybe you can try efficient net and also oversampling the dataset might will help,thanks
If you’re working on ultrasound images, is it correct to assume you are trying to perform segmentation? If so, you should try U-net as it has been widely used in medical image processing. MaskRCNN works for general objects like cars, human beings, etc. Unfortunately, no amount of data augmentation or changing backbone will help in your case. They will provide at most very little boost; which won’t be enough.
Heres the first paper for UNet which is popularly used to segment medical images:
Article U-Net: Convolutional Networks for Biomedical Image Segmentation
heres the keras implementation:
https://github.com/zhixuhao/unet
and heres an improved version:
https://arxiv.org/pdf/1902.04049.pdf
Dear Syed Mobassir Hossen and Tahmid Chowdhury Thank you for your answers.
The MaskRCNN algorithm allows you to use any pre-trained weights and a variety of backbones, or even use your own. There's a great tutorial where it was used to identify nuclei.
The real question is whether or not you need the polygon maks, If you don't, then you are fine with UNET for segementation, but if you also want your model to return a polygon mask outlining each feature (for positional information), then you will need something like MaskRCNN.
This paper was pretty helpful:
https://arxiv.org/pdf/1901.10170.pdf
Also, for MaskRCNN, there are sample datasets that come with their own Jupyter Notebooks that you can use to evaluate the effect of each parameter you change in while training. I would second the recommendation to use resnet101 and supplement your training set with publicly available data.
If your network trained successfully, then it is Ok. But is it trained successfully? What were the values of loss and val_loss during training?
If it was overfitting,
You have to add more data, apply augmentation, hyper parameter tuning, add drop out, and L2 regularization to avoid overfitting.
For segmentation tasks, instead of Mask-RCNN, you can use U-Net, LinkNet or PSPNet. You can also use different feature extractor at the down-sampling part of your architecture, i.e. InceptionV3, Xception, DenseNet121.
Hey Summer Rae I am using Maskrcnn for my thesis work. However, I am not able to train it on all layers when using resnet101 backbone.
The most probable reason for that is large image size 1600,1600. Do you have any method in your knowledge to overcome this problem??
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training