EMC-Mini-project

---

Problem → Idea → Outcome

• Problem: Simulating a full EMC dictionary over many T2 values is slow (1-4 hours).
• Idea: Skip T2 points where the EMC curves change slowly, keep dense sampling where they change fast, then interpolate along T2.
• Outcome: Same fitting pipeline, faster dictionary generation. In our test: MAPE ≈ 0.01% vs. the full dictionary.

---

Why the decay is not mono-exponential

Multi-spin-echo (TSE/CPMG) trains don’t follow a simple S(t)=S0·exp(−t/T2) (as in theory) because:

• Imperfect refocusing (B1⁺ inhomogeneity / flip-angle errors) creates stimulated echoes that mix with primary echoes.
• Slice-profile & crusher schemes redistribute coherence pathways (EPG formalism), altering the apparent decay.
• Optional effects (exchange, diffusion during gradients) further deviate from a single exponential.

➡️ Therefore we use a Bloch/EPG-simulated EMC dictionary over (T2, B1⁺, …) rather than fitting a mono-exponential.

---

How we measure success (comparison factors)

We compare the T2 map from the full dictionary to the T2 map from the interpolated dictionary.
Per-voxel percent error is: err(i) = 100 * ( T2_interp(i) - T2_full(i) ) / T2_full(i) % (Computed only on valid voxels: finite and > 0. DICOM values are first rescaled using RescaleSlope/Intercept.)

Reported metrics

• Pixels compared — number of voxels.
• Bias (mean %) — average signed % error. Goal: ≈ 0%.
• MAPE (mean abs %) — average abs % error. Goal: ≪ 1%.
• Median |%err| — robust central tendency. Goal: ≈ 0%.
• 95th pct |%err| — tail error; 95% of voxels are below this value. Goal: < ~1%.

---

Results (snapshots)

Left: T2 map with the full dictionary vs. with the interpolated dictionary.
Right: Voxel-wise % error (interp vs full), display range ±1%.

Example stats:

• Pixels compared: 9,177
• Bias: 0.00% MAPE: 0.01% Median |%err|: 0.00% 95th pct |%err|: 0.00%
• Dictionary diff (relative Frobenius): 8.7e-4

---

How to run Task2

1. Setup

   • Clone this repo and open MATLAB.
   • Add sources to the path:

     addpath(genpath('src'));

   • Point dicPath to your EMC dictionary (.mat that contains T2_tse_arr and echo_train_modulation).
     Example:

     dicPath = 'data/SEMC149.mat';  % or an absolute path on your machine

2. Create an interpolated dictionary (Task 2)

   • Define the banded rules (ms) and run:

     rules = struct( ...
       't2_min',{  1,  81, 301}, ...
       't2_max',{ 80, 300, Inf}, ...
       'stride',{  5,  10,  25});
     
     [interpDicPath, pctSaved, keepMask] = task2_adv_interp(dicPath, rules, 'pchip');

   • This writes *_task2_interp.mat next to your source dictionary and prints the % time saved.

3. Generate T2 maps in the GUI

   • Open the EMC T2 FIT GUI (per your setup).
   • Load the experimental dataset.
   • Run baseline with the full dictionary (B1 fit OFF, per assignment) and export the DICOM.
   • change file name to 'T2_map_Full'.
   • Run again with the interpolated dictionary (interpDicPath) and export the DICOM (e.g., ...T2map_EMCt-task2_interp.dcm).
   • change file name to T2_map_interpAdvanced.

4. Compare maps (voxel-wise % error)

   compare_emc_t2_dicoms( ...
     'path/T2_map_Full.dcm', ...
     'path/T2_map_interpAdvanced.dcm');