Segmentationsummary
Initial Discussion with Sudhir regarding Segmentation
When a tumor is present inside a DICOM dataset, segmentation helps us separate that tumor by outlining it from the rest of the tissues so we can look at it more clearly. We discussed about to start with 2D segmentation, where we look at each slice of the scan one by one (like flipping through pages of a book) and outline the area where the tumor is present. The process typically starts with 2D segmentation, where each DICOM slice (axial, coronal, or sagittal) is analyzed individually to detect and outline the tumor boundaries. This can be achieved AI-based algorithms to create accurate 2D segmentations for slices At this stage, the complex thing would be is to build a complex polygonal mesh/object that represents the tumor shape in 3D or use a “shrink-wrap” method that wraps a smooth outer layer around the segmented region to form a realistic 3D model.
1. Brain Tumor Segmentation
Overview
The brain tumor segmentation work is based on a U-Net deep learning approach applied to MRI scans from the MICCAI BraTS dataset. The objective is to automatically segment tumor regions by learning from multi-modal MRI inputs (T1, T1c, T2, and FLAIR).
The notebook demonstrates an end-to-end pipeline that includes data preprocessing, model training, and prediction of segmentation masks. MRI slices are processed and fed into a U-Net architecture, which learns to distinguish tumor regions from healthy tissue. The model outputs pixel-wise segmentation masks that highlight tumor areas, enabling accurate localization and analysis.
This approach emphasizes supervised learning using labeled medical imaging data and showcases how convolutional neural networks can effectively perform biomedical image segmentation tasks.
Dataset Used
MICCAI BraTS 2019 Dataset
Multi-modal MRI dataset including:
T1
T1c (contrast-enhanced)
T2
FLAIR
Widely used benchmark dataset for brain tumor segmentation
Provides labeled ground truth masks for tumor regions
Notebooks and Models Used
1. Brain Tumor Segmentation U-Net (Zeeshan's Original)
- Base Architecture: U-Net
- Dataset: MICCAI BraTS 2019
- Platform: Kaggle Notebook
- Runtime: ~4 hours (CPU), ~1h 35m (L4 GPU)
Kaggle Notebook link: https://www.kaggle.com/code/zeeshanlatif/brain-tumor-segmentation-using-u-net
What was done:
- Implemented a U-Net based deep learning model for tumor segmentation
- Processed MRI slices and trained the model to predict segmentation masks
- Generated 2D segmentation outputs for each slice
- Evaluated segmentation quality using dataset-provided labels
2. Nasopharyngeal Carcinoma Segmentation
Overview
This work focused on segmenting nasopharyngeal carcinoma (NPC) tumors from MRI datasets. Unlike the BraTS dataset, this dataset required additional preprocessing and restructuring before it could be used for model training.
The approach reused the U-Net architecture from the brain tumor segmentation work, adapting it to a different anatomical region and dataset format. Since the dataset was not originally structured like BraTS, significant effort was spent on converting and organizing the data to ensure compatibility with the existing training pipeline.
Dataset Used
Primary Nasopharyngeal Carcinoma MRI Dataset
MRI-based dataset focused on NPC tumor regions
Data was converted from original format to NIfTI (.nii)
Dataset was restructured to align with BraTS-like format for compatibility
Notebook and Model Used
Nasopharyngeal Carcinoma Segmentation Model
- Base Architecture: U-Net (Zeeshan's model as backbone)
- Dataset: Primary Nasopharyngeal Carcinoma MRI (converted to NIfTI)
- Notebook: Custom notebook (dataset restructured to match BraTS format)
What was done:
- Converted raw MRI data into NIfTI format for processing
- Restructured dataset to follow BraTS-style directory and labeling format
- Reused the existing U-Net pipeline designed for brain tumor segmentation
- Trained the model on NPC dataset for tumor region detection
- Generated slice-wise segmentation masks for nasopharyngeal tumors
3. Sinus Segmentation
Dataset Used
NasalSeg Dataset (Sinus CT Data)
CT-based dataset focused on sinus and nasal cavity regions
Input: CT volumes (.nii format)
Output: Corresponding segmentation masks (.nii format)
Requires paired data:
CT images
Ground-truth masks (same filenames)
Dataset access was requested via email (no response received)
Repository and Model Used
SinusSegment Model
- Base Architecture: UNet++
- Dataset: NasalSeg Dataset
- Repository: https://github.com/rheadkaul/SinusSegment
What was done
- Used the SinusSegment repository for sinus segmentation workflow
- Trained a UNet++ model using CT scan volumes and corresponding masks
- Performed inference on test CT scans to generate predicted segmentation masks
- Evaluated model performance using segmentation metrics
- Saved output masks in NIfTI format for further visualization and analysis
nnInteractive Research Paper Datasets
Format: Various
Status: Listed in research paper
Source: GQC SharePoint
Explored as part of understanding interactive segmentation approaches and available datasets used in research
4. Research Papers Referenced
Overview
Several research papers were reviewed to understand segmentation techniques, model architectures, and dataset usage. These papers guided the selection of models such as U-Net and UNet++, and provided insights into both automatic and interactive segmentation approaches.
5. Tools Used
Overview
Various tools and documentation resources were used to support segmentation workflows, including data visualization, annotation, and deep learning-based segmentation utilities.
Tools Used
ITK-SNAP Software
Type: Software Tool
Used for:
Visualizing medical imaging data (DICOM, NIfTI)
Manual and semi-automatic segmentation
Inspecting segmentation outputs and masks
Link: https://www.itksnap.org/pmwiki/pmwiki.php?n=Downloads.SNAP3
ITK-SNAP DLS (Deep Learning Segmentation) Documentation
Type: Documentation
Used for:
Understanding deep learning-based segmentation workflows
Running pre-trained models and inference pipelines
Exploring interactive segmentation capabilities