The dataset consists of radiological images, X rays, from over 1000 thousand patients who visited the Indian Spinal Injuries Centre seeking for their treatment. The X-ray images of the patients focus or highlight the thoracic and lumbar regions of the vertebral column. For each patient in the dataset, there are two X-Ray images highlighting two different views namely AP and Lateral, that help in identifying the underlying abnormality vividly. This gives us a total of 2000 X-ray images, equally divided among the two views.
Here we have performed two tasks A) Classification and B) Segmentation
For classification task, we had to identify patients with damaged or normal spinal cord according to the characteristics of their AP and LAT images. In this case we take the original RGB image for both LAT and AP views and down sample them to 224x224x1. Then augment AP and LAT images together and input it to the neural network. The output of the network will be a label (Damaged or Normal).
We adjusted the validation split in data to 0.1 ratio. Accuracy was used as metric and the validation acc was 85%.
Conv2d_1_input:InputLayer
conv2d_1:Conv2D
batch_normalization_1()
activation_1()
max_pooling_2d_1
conv2d_2:Conv2D()
batch_normalization_2()
activation_2()
max_pooling_2d_2()
conv2d_3:Conv2D()
batch_normalization_3()
activation_3()
max_pooling_2d_3
dropout_1()
flatten_1()
dense_1()
batch_normalization_4()
activation_4()
dense_2
For the segmentation task, we had to create 3 masks for the AP image and 5 masks for the LAT image. In this case we take the original RGB image for the two cases separately, down sample it to 224x224x3 and then input it to the neural network. The output of the network will be 224x224 masks (single-channel). Hence we trained eight different models. Some test masks are shown below:
We use the UNET Architecture. The architecture is one of the state of the art architectures for bio-medical image segmentation. Our U-Net implementation has 4 downblocks, 1 bottleneck layer and 4 upblocks.