Master-thesis : Comparative Evaluation of Deep Learning Networks for Intracranial Aneurysm Segmentation

This repository contains the code and data processing pipeline for the master thesis project.
The project involves the preparation, training and evaluation of the nn-UNet segmentation model

---

Table of Contents

1. Requirements
2. Data Exploration
3. Data Preparation
4. Total Segmentator
5. NN-UNet pipeline
6. Evaluation
7. References

Requirements

• Python 3.12.4
• Required Python packages (list in requirements.txt)
• nnU-Net version 2.2.

Data Exploration

• Check_image_label_shapes checks that each image has a corresponding label of the same shape. For the CT dataset one of the images didn't have a label of the same size, so it was discarded.
• Checking_image_resolution categorizes images from the CT dataset into groups of different resolution (in terms of voxel size).
• count_an_per_size_trainset counts aneurysm sizes in the original train dataset in different size categories, with my definition of aneurysm size.

Data Preparation

• To use NN-UNet, generate_dataset_json_file and dataset_conversion_file were used to create the expected folder structure and to follow the NN-UNet naming conventions.

CT dataset (id = 059)

Flowchart for the CT data preparation and nnunet commands ( without TotalSegmentator )

• CT raw dataset was collected from open source Zendo ( see reference), including training images (imagesTr, n = 1186)) and test images (imagesTs_internal, n= 152)).
• The whole dataset was copied to perform further modifications on the copy ( renaming, preprocessing etc).
• Rename data to nnU-Net format and store old and new names in a lookup table, running nnunet commands like verify data integrity, preprocessing and train. All these steps are described here.
• Rename the test set too ( it was named differently from the train data).

Flowchart fro the CT data processing with Total Segmentator

MR dataset (id = 060)

• Originally, in the MR dataset available online, only patients with aneurysms had corresponding label files, the control participants didn't have label files. I created empty images with the same dimensions to provide nnunet with images and labels files in pairs.
• For patients with multiple aneurysms, there were multiple label files with one aneurysm per each label file. I merged multiple lesion in the same label files so each aptient has only one corresponding label.
• Files should have been renamed to nnunet format.
• Split all training files into validation and test here.
• The MR images had been pre-cropped to encompass the brain and the specific region of interest within the brain (Circle of Willis). Consequently, the MR dataset did not require processing through the Total Segmentator step, unlike the CT dataset.

Total Segmentator

• See instructions to use TotalSegmentator here.
• The bash script to run TS over multiple images in a folder
• Once the brain masks are available, we crop the original images to the brain region and add padding.
• Warning: check the volume of the label before and after cropping, TS shouldn't cut out labels inside the brain. But it is possible to see aneurysms outside of the brain.

NN-UNet pipeline

NN-UNet training consists of runnnig the following commands consecutively :

1. nnUNetv2_plan_and_preprocess and --verify_dataset_integrity - Generate dataset fingerprint and plans
2. nnUNetv2_train - Train the models using the 5-fold cross-validation approach, choose 3d_fullres
3. nnUNetv2_find_best_configuration- Find the best configuration for the dataset using the specified folds.
4. nnUNetv2_predict - Generate predictions for the internal test set.
5. nnUNetv2_apply_postprocessing - Apply post-processing to the predictions to refine results.
6. nnUNetv2_evaluate_folder - Evaluate postprocessed perdictions compared to the ground truth labels.

This bash script runs all these steps automatically from step 1 to 3. Steps from 3 till 6 are automated [here](NN-Unet pipeline/script_060_4.sh). There's a break at step 3 because it needs user intervention. You have to see generated instructions and run the code as specified.

The post-processing is done by the default nnunet procedure, the code can be found here, remove all but the largest component code is here. The way it works is that it first does a connected component analysis, which also gives the size per component and then filters for the largest component. An additional post-processing step is removing connected components smaller than 1mm in diameter. This is done in the evaluation script (see below).

Evaluation

NN-UNet generates an evaluation summary that gives some basic metrics across all images on the vowel-level. I implemented my own evaluation pipeline automated here.

References

• nnUnet : Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, https://doi.org/10.1038/s41592-020-01008-z .

• TotalSegmentaor : Akinci D'Antonoli, T., Berger, L. K., Indrakanti, A. K., Vishwanathan, N., Weiß, J., Jung, M., Berkarda, Z., Rau, A., Reisert, M., Küstner, T., Walter, A., Merkle, E. M., Segeroth, M., Cyriac, J., Yang, S., & Wasserthal, J. (2024). TotalSegmentator MRI: Sequence-Independent Segmentation of 59 Anatomical Structures in MR Images. arXiv preprint arXiv:2405.19492. https://doi.org/10.48550/arXiv.2405.19492.

• CT dataset : Z.-H. Bo. Large IA Segmentation dataset. Version 1. This dataset is for non-commercial purposes. Feb. 2021. DOI: 10 . 5281 / zenodo . 6801398. URL: https://zenodo.org/record/6801398 .

• MR dataset : T. Di Noto et al. “Towards Automated Brain Aneurysm Detection in TOFMRA: Open Data, Weak Labels, and Anatomical Knowledge”. In: Neuroinformatics 21.1 (2023), pp. 21–34. ISSN: 1559-0089. DOI: 10.1007/s12021-022- 09597-0. URL: https://doi.org/10.1007/s12021-022-09597-0.