CT Analysis with SaxNeRF

Overview

This repository documents 3D reconstruction experiments on multi-wavelength CT datasets from Kubota et al. The datasets contain compositions of materials with varying densities (polymer, metal, etc.), scanned at 3 different wavelengths to produce distinct absorption projections.

The detector used is a Carbon Nano Tube (CNT) detector, capable of detecting wavelengths across an ultra-wide range. For full details on the acquisition setup, refer to the Kubota et al. paper.

The goal is volumetric 3D reconstruction. While the original paper used shape-from-silhouette and visual hull reconstruction, this project explores SaxNeRF as an alternative.

---

Dataset

Raw datasets are located at:

/home/yixin/Desktop/NII佐藤先生-20251107T091146Z-1-001

The source computer is at desk 7, room 1612.

The detector has a resolution of 3 × 20 pixels (height=3, width=20).

Inside （久保田君）CTイメージングの多波長データ/1,310 nm, there are two input files:

File	Description
onoff.csv	Per-pixel detector intensity recorded over ~100 timesteps (illuminated → dark)
CTdata.csv	Projection data, one row per acquisition angle. Each row contains 60 comma-separated values to be reshaped into a 3×20 matrix

---

Data Preprocessing

The preprocessing pipeline is defined in calibration_1310.ipynb. It converts raw string/CSV data into calibrated projection intensities.

Key steps:

1. Parse and calibrate raw intensities from onoff.csv and CTdata.csv
2. Scale down projections by a factor of ~0.07 (empirically determined for SaxNeRF convergence)
3. Export projections to pickle format: cnt_360.pickle

---

Training

The SaxNeRF repository is located on NII server 2201NII:

/home/yixin/workspace/SAX-NeRF

Setup:

1. Place the generated cnt_360.pickle inside the data/ folder
2. Create a training config at config/Lineformer/cnt_360.yaml

Start training:

train_mlg --config config/Lineformer/cnt_360.yaml --gpu_id 0

Monitor training:

logs/Lineformer/

---

Results

Full results are documented in the CNT_Experiments presentations.

Reconstruction from CNT Projections

SaxNeRF fails to converge on the original Kubota-san dataset. The primary cause is the very low vertical resolution of the projections (only 3 pixels tall). Approaches such as repeating or stretching the volume along the vertical axis did not yield meaningful improvements.

Toy Volume Experiments

To isolate the failure mode, a synthetic toy volume was created — a cuboid with a square hole — and projections were generated using the TIGRE toolkit (installed inside the SaxNeRF repo).

Pipeline for generating projections from a 3D voxel volume:

volume.npy  ──►  dataGenerator/raw_data/cuboid/volume.npy
                 dataGenerator/raw_data/cuboid/config.yml   (projection settings)
                         │
                         ▼
              data_vis_cuboid.py       (intermediate .mat file + visualizations)
                         │
                         ▼
              generateData_cuboid.py   (outputs pickle to data/)
                         │
                         ▼
                   data/cuboid.pickle  ──►  SaxNeRF

Convergence Limits

Through systematic experiments with the toy volume, the following thresholds were identified:

• ✅ SaxNeRF can reconstruct volumes from as few as ~10 projections
• ⚠️ A minimum of ~20 pixels on both spatial dimensions is required for a decent reconstruction
• ❌ The original CNT dataset (3-pixel vertical resolution) falls below this threshold

---

Repository Structure

SAX-NeRF/
├── config/
│   └── Lineformer/
│       └── cnt_360.yaml          # Training hyperparameters
├── data/
│   └── cnt_360.pickle            # Preprocessed projections
├── dataGenerator/
│   ├── raw_data/cuboid/
│   │   ├── volume.npy
│   │   └── config.yml
│   ├── data_vis_cuboid.py
│   └── generateData_cuboid.py
└── logs/
    └── Lineformer/               # Training logs

---

References

• Kubota et al. — Multi-wavelength CT imaging with CNT detector (see paper for full citation)
• SAX-NeRF — Sparse-view CT reconstruction with NeRF
• TIGRE Toolkit — Tomographic iterative GPU-based reconstruction