kaggle_mgmt_status

Attempt at Kaggle mgmt competition with Raghav Even though there is no evidence that there is any correlation between mri brain scans and presence/absence of mgmt status
It was a good learning experience

Methods

1. Conv3Dnet
   • Since mgmt status can be determined in scans with tumor.we first filter out data with only tumor
   • We trained a simple cnn model for tumor classification, this reduces the depth and allows the model to focus on important slices
   • trained a resnet18_3d/efficientnet_3d/vanilla_3d with variable depth
   • Model overfitting with 100% acc on train dataset,performing poorly on validation dataset
2. Pretrained Conv2d + transformer
   • Features and deep layer embeddings of a image stacked with all slices fed into a time series model like transformer
   • Batch_size,height,width,depth,channels-> Conv2d -> Batch_size,depth,features -> transformer -> classification
   • Not enough information in feature space
3. voxel + Conv3dnet
   • Bring all patients scan into 1 modality using voxel orientation code
   • Use the same pipeline as 1