Digits not Classifying

I am using Digits 6.1.1 with Caffe version 0.15.14 and flavor NVIDIA. I am NOT using docker.
I am trying to use Digits to classify a healthy brain with an Alzheimer’s Disease brain but Digits is not classifying well as the validation and test each have an accuracy around 50-60%

I squashed the images which were 91x109 to 28x28 gray scale PNG files. I have 25 brains for each class. Both the validation and training images are the same size and are gray scale.
I am using the Lenet model with images that are 28x28 gray scale (which is recommended).
I have tried subtracting the mean image and no subtractions.
I am using 30 epochs with SGD and a base learning rate of 0.01 with a step down at a step size of 33% and a Gamma of 0.1

What can we do to increase the accuracy? (The logs are below)
caffe_output.log (111 KB)
caffe_output.log (111 KB)

There could be so many reasons why this is not working.
How many images do you have per class?
Maybe your images are to small. Ask yourself… would you be able to classify these 28x28 small images?
LeNet is a very small net… maybe you are better off with an more capable network.

Your logs are showing that your network is overfitting. After epoch 2 your validation performance is decreasing from 0.72 towards 0.61.

I have tried using GoogLeNet and AlexNet and all give me the same result. I have changed the image sizes and still get the same result. I have 25 training subjects for each class each with 2 MRI sequences (each MRI has about 80 slices).

Here is a trial with AlexNet without squashing the image.
caffe_output.log (133 KB)

Here are the logs and a sample of the Concatenated images. I used GoogLeNet and LeNet but did not get the results I hoped for
Is there something wrong with my data? Is it better to concatenate or keep each slice as its own image? Do i need to customize the model- if so how and what do I need to do?
caffe_output.log (45.1 KB)
caffe_output.log (356 KB)


What do you expect? 100%? Your LeNet achives ~81% after the 1 Epoch. AlexNet ~ 75% after 1 Epoch.

  1. You probably don’t have enough data
  2. Your data might be to ambiguous; Which of your slices show damaged area of the brain
  3. start reading how CNNs work
  4. start reading the literature e.g. [1906.04231v1] Alzheimer's Disease Brain MRI Classification: Challenges and Insights

I was hoping for something a bit more like 85% since a study using digits received almost 100%
I have tried with 90 patients for each class (180 scans for each class) and still received the same results.
I wanted to include the whole brain since I was doing a man bs machine competition.
I tried copying this study but with structural MRIs https://devblogs.nvidia.com/nvidia-digits-alzheimers-disease-prediction/