Need suggestions : Classification of Medical images using Caffe. Am i right?

[ChanChip]

Hello all,

  My objective is "Classiifcation on medical images using deep conv. networks"

   I'm quite puzzle about the results which I obtained from classification of medical image of my own dataset after a long run on CPU not GPU. In details, I have 50 medical images. I split them in to two sets 1. training set(25 images with labels) 2.Validation set(25 images with labels). I chose caffe framework for this process.

  Step 1 :  created own lmdb database with that training and validation images

  Step 2 :  Modified train_*.sh, *_solver.prototxt, *_train_test.sh(learning rate, switch to CPU, Layers(conv, Relu,Loss layer added), 10000 iterations)

  Step 3 : ran train_*.sh

In CPU mode, it ran around 35 hours for completion.

My doubts are follows :

    1. Starting onwards(0th to 10000) it shows the Testing accuracy as 24%(no fluctuation at any point)

    2. Training loss getting change but Testing loss keep steady and very little changes.

  Appreciate that any advice/helps from experts.

Thanks,

DeepLover

      

[Neil Nelson]

I just downloaded the PASCAL VOC 2012 dev kit and see that there are 17125 images. There are 24 classifications (different labels) across that number. That gives around 714 images per label. That is not quite correct in that there are usually several labels per image. Having a good number of images per label will be required to get reasonable accuracy. This observation may not address the reported problem but will certainly be important overall.

[ChanChip]

  First of all, thanks Neil. 

  "Having a good number of images per label will be required to get reasonable accuracy" - That's correct.So, it means that "Images with exact labels E.g Cat image with cat label yield high accuracy"(this is my understanding). 

    If you don't mind How can i do the same for medical images? should I have to label the images as human observer point of view e.g. classified as normal, mild, severe,moderate ? or any special algorithms is there for labeling those images? 

   Earlier,I've used some algos. and obtained those labels for 25 training images and 25 validation images. It brought me just 24% accuracy results.

Thanks,

Deeplover