This Jupyter file is my submission for the RSNA Spine Challenge
Method-1 (recommmended): Run it directly on Kaggle. You can find my notebook here, with all the necessary datasets attached
Method-2: Run it locally. The original dataset is approximately 30 GB, but I processed it down to 4.5GB. You can download the dataset, and change the file paths, to run the notebook locally.
load_dicom_stack(): Processes a single MRI image into a Numpy array
create_dataset(): Parses 30GB of MRI images to create a 4.5GB Numpy dataset. The function took 8 hours to run, and I saved the output, so you won't have to rerun the function
create_tf_dataset(): Labels the dataset and formats it as a TensorFlow dataset, ready for model input.
complete_training_workflow(): creates and trains the model
complete_testing_workflow(): runs the trained model on the test set and converts the the model prediction (a Numpy array) to a submission.csv file
Implement image segmentation using Medical SAM2
Use the DenseNet architecture instead of the EfficientNet, as this article suggests that older architectures perform better on MRI images
Explore domain adaptation techniques to improve the performance, because the images used in ImageNet are significantly different from MRI images
Initially, I tried to use the DeepSpine model architecture. Its a multi-input, multi-task, and multi-class model with SOTA performance. The model weights were not available online, so I decided to create and train the 50-million-parameter model from scratch. Unfortunately, I ran out of RAM while doing this.
Then, I decided to use transfer learning with a pre-trained EfficientNetV2 model. The problem? The model expects image of dimension x * x * 3 while my images were of the dimension x * x * 60 * 3. An MRI is a stack of grey-scale images, just like a video is a stack of RGB images. Its unlike the images present in the ImageNet dataset. The articles how to get started with deep learning using MRI data and dealing with MRI and Deep Learning with Python saved my life.
The hardest part was figuring out how to process the images, labelling the dataset, and formatting it so that it can be fed into the model—all while keeping the memory usage under 30GB of RAM
After dedicating over two months to this project, I finally submitted my prediction. The result? I was ranked 954/1010. I am pleased to be in the 10%, nevermind it's the bottom 10% instead of the top