Merge pull request #14 from ChristianHinge/main

README links, updates and polish

Author: ChristianHinge

README.md

@@ -8,7 +8,7 @@
 
 ---
 ## Updates
-
+- April 8, 2026 (13:00 GMT+2): [BUG FIX] A bug in the Carney HU-to-MU conversion formula has been fixed. Please make sure to run `docker pull ghcr.io/bic-mac-challenge/recon:latest` and `git pull` for updated recon and evaluation logic. The 8 ground truth PET train images have been updated on huggingface. 
 - April 7, 2026: [NEW DATA POLICY] The use of public datasets for pretraining and other use-cases is now allowed under certain conditions. Please see [docs/rules.md](docs/rules.md) for details. 
 ## Table of Contents
 
@@ -188,7 +188,7 @@ Five metrics compare predicted PET and CT outputs against the ground truth:
 | Metric | Modality | Description | Region |
 |--------|------|-------------|--------|
 | Whole-body SUV MAE | `PET` | Mean absolute error in standardised uptake value (SUV = activity × weight / total dose) | Body mask, excluding ±4 cm around liver |
-| Brain Outlier Score | `PET` | AUC of fraction of brain voxels within relative error thresholds (5%, 10%, 15%) | Brain |
+| Brain Outlier Score | `PET` | AUC of fraction of brain voxels within relative error thresholds (5%, 10%, 15%) [Ladefoged 2017 (fig. 8)](https://www.sciencedirect.com/science/article/pii/S1053811916307170)| Brain |
 | Organ Bias | `PET` | Mean absolute relative error of mean SUV in 8 organs: brain, liver, spleen, heart, pancreas, muscle, adipose, extremities | TotalSegmentator organ labels |
 | CT MU MAE | `CT` | Mean absolute error of attenuation coefficients (μ at 511 keV) between predicted and ground-truth CT after HU→μ conversion | Body mask, excluding ±4 cm axial slices at top of liver|
 | TAC Bias | `Dynamic PET` | Absolute relative error of the integral of time-activity-curves (TACs) for the aorta and selected brain regions. NOTE: Metric is computed only for the final test set due to the size of the dynamic sinograms. | Brain regions and aorta|

docs/data-background.md

@@ -22,7 +22,7 @@ A typical session ran as follows:
 
 ## Feature modalities and alignment
 
-The four input modalities span a spectrum from *well-aligned but low-quality* to *high-quality but poorly aligned* relative to the ground-truth CT:
+The three input modalities span a spectrum from *well-aligned but low-quality* to *high-quality but poorly aligned* relative to the ground-truth CT. All images under `features/` have been resampled to the CT grid (512×512×531). The topogram is 2D and therefore only resampled in the frontal plane (512×1×531).
 
 ### Topogram (`topogram.nii.gz`)
 A 2D sagittal scout X-ray acquired roughly 2 minutes before the CT. It is the most temporally and spatially aligned feature, but has very limited anatomical detail — it is a single 2D projection with coarse resolution. The topogram and CT share the same bed position and scanner table.
@@ -33,6 +33,8 @@ The non-attenuation-corrected PET, reconstructed from the **50–70 minute** win
 ### MRI (`mri_chunk_*` / `mri_combined_*`)
 Whole-body DIXON MRI acquired on a **separate scanner** the same day as the PET/CT (for 95/99 participants; 4 were scanned 9–89 days later). The MRI provides excellent soft-tissue contrast and anatomical detail, but it is not inherently aligned to PET/CT space. The DIXON MRI sequence used is designed specifically to capture anatomical informing relevant for attenuation correction. **The whole-body image was acquired in four sequential scans of 25s coressponding to four body chunks**. Prior to each acquisition the participant was asked to take a deep breath and hold still. Since participants were not asked to hold their breath during Topogram, CT or PET acquisition, the lung volume is often larger on the MRI. Participants were placed head first supine with arms down the side in both PET/CT and MRI, but the use of coils, different beds and raisable head rests makes the MRI-to-PET/CT alignment poor. A rigid translation (no rotation) was applied to bring the MRI into approximate PET/CT alignment. Whether to incorporate additional registration as a preprocessing step is left to the participant.
 
+
+
 ### Summary
 
 | Feature | Alignment to CT | Anatomical quality |
@@ -41,5 +43,4 @@ Whole-body DIXON MRI acquired on a **separate scanner** the same day as the PET/
 | NAC-PET | Good (same scanner, ~50 min later) | Moderate |
 | MRI | Approximate (rigid translation only) | High |
 
-> [!NOTE]
-> All features under `features/` have been resampled to the CT grid (512×512×531). The topogram is 2D and stored as (512×1×531). Despite sharing the same matrix, this does not mean they are perfectly registered — resampling preserves the grid, not the alignment.
+ Despite all images sharing the same matrix, this does not mean they are perfectly registered.

docs/reconstruction.md

@@ -115,4 +115,5 @@ Only subjects in the full `train/` split and the `val/` split include `recon/` d
 ## Further Reading
 
 - Carney et al. (2006) — *"Method for Transforming CT Images for Attenuation Correction in PET/CT Scanners"*, Medical Physics. The bilinear HU→μ model used in this pipeline.
-- Thielemans et al. (2012) — *"STIR: Software for Tomographic Image Reconstruction Release 2"*, Physics in Medicine and Biology. The reconstruction library used here.
+- Cherry, Sorenson & Phelps — *Physics in Nuclear Medicine* (4th ed., 2012)
+