Skip to content

maxmelichov/RSNA-2024-Lumbar-Spine-Degenerative-Classification

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

42 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

RSNA-2024-Lumbar-Spine-Degenerative-Classification

What we need to do: Model output is 3 multilabel classes. class balance

you have 26 spots with the problem of the classification spot 1-3 spot 4-6

Trials:

Trial 1:

using davit_small with hot_category with Sagittal_T1, Sagittal_T2, Axial_T2 on top of each other image size 384x384x10 for Axial_T2 and 192x192x15 for Sagittal_T1/T2 CustomLoss severeloss using timm/davit_small.msft_in1k net number of epochs: 5 results: fold 1: Train Loss: 0.802 Validation Loss: 0.806 Last LR [0.000036] Kaggle score: 1.08

Trial 2:

using davit_small with hot_category with Sagittal_T1, Sagittal_T2, Axial_T2 on top of each other image size 384x384x10 for Axial_T2 and 192x192x15 for Sagittal_T1/T2 crossentropy with weights using timm/davit_small.msft_in1k net number of epochs: 5 results: fold 1: Train Loss: 0.792 Validation Loss: 0.728 Last LR [0.00036] Kaggle score: 1.03

Trial 3:

using davit_small with hot_category with Sagittal_T1, Sagittal_T2, Axial_T2 on top of each other image size 384x384x10 for Axial_T2 384x384x15 for Sagittal_T1 and 192x192x15 for Sagittal_T2 crossentropy with weights using timm/davit_small.msft_in1k net and for feature extactor timm/efficientnet_b3.ra2_in1k number of epochs: 5 results: fold 1: Train Loss: 0.76 Validation Loss: 0.79 Last LR [0.00036] Kaggle score: 0.98

Trial 4:

using davit_small with hot_category with Sagittal_T1, Sagittal_T2, Axial_T2 on top of each other image size 384x384x10 for Axial_T2 384x384x15 for Sagittal_T1 and 192x192x15 for Sagittal_T2 crossentropy with weights using timm/davit_small.msft_in1k net and for feature extactor timm/efficientnet_b3.ra2_in1k using 5 models for each part number of epochs: 5 results: fold 1: Train Loss: 0.749 Validation Loss: 0.585 Last LR [0.00036] Kaggle score: 0.95

Trial 4:

using davit_small with hot_category with Sagittal_T1, Sagittal_T2, Axial_T2 on top of each other image size 384x384x10 for Axial_T2 384x384x15 for Sagittal_T1 and 192x192x15 for Sagittal_T2 crossentropy with weights using timm/davit_small.msft_in1k net and for feature extactor timm/efficientnet_b3.ra2_in1k using 5 models for each part number of epochs: 10 results: fold 1: after 4 fold the results on kaggle are the same Train Loss: 0.749 Validation Loss: 0.585 Last LR [0.00036] Kaggle score: 0.95

Trial 5:

Summary:

The model is designed to take in axial and sagittal MRI scans of different sections of the lumbar spine, extract features using powerful pre-trained deep learning models (ConvNeXt for sagittal and axial), and then combine these features to predict some region-specific outcomes. The use of separate backbones for different regions and the combination of axial and sagittal features makes the model highly specialized for analyzing lumbar spine MRI data.

image size 512x512x3 for Axial_T2 128x128x5 for Sagittal_T2 crossentropy with weights using timm/convnext_nano.in12k number of epochs: 5 results: fold 1: after 4 fold the results on kaggle are the same Train Loss: 0.6 Validation Loss: 0.7 Last LR [0.00036] Kaggle score: 0.8

Trial 6:

Summary:

The model is designed to take in axial and sagittal MRI scans of different sections of the lumbar spine, extract features using powerful pre-trained deep learning models (ViT for axial and ConvNeXt for sagittal), and then combine these features to predict some region-specific outcomes. The use of separate backbones for different regions and the combination of axial and sagittal features makes the model highly specialized for analyzing lumbar spine MRI data.

image size 518x518x3 for Axial_T2 128x128x5 for Sagittal_T2 not using Sagittal_T1 crossentropy with weights using timm/convnext_nano.in12k for Sagittal_T2 and timm/vit_small_patch14_reg4_dinov2.lvd142m and for Axial_T2 number of epochs: 5 results: fold 2: after 4 fold the results on kaggle the result is much worst 1.36 Train Loss: 0.48 Validation Loss: 0.47 Last LR [0.00036] Kaggle score: 0.64 for fold 1 I got 0.62

Trial 7:

Summary:

In this trial, a single model was used sequentially for different types of images: Sagittal_T1 first, followed by Sagittal_T2, and finally the last three images of Axial_T2. The preprocessing involved cutting the appropriate spinal levels for the Sagittal images and extracting the central part of the Axial images to ensure a consistent size of 152x152 pixels. The images were processed with a total of 29 color channels. Model used: timm/edgenext_base.in21k_ft_in1k Number of epochs: 5 Results: Training Loss: 0.52 Validation Loss: 0.57 Last Learning Rate: 0.00036 Kaggle Score: 0.55

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published