Fix: deg/rad unit mismatch in removeProjections corrupts chunked helical FBP

[zezisme]

Bug Summary

In helical cone-beam FBP, when the projection + volume data exceed GPU memory, LEAP triggers backproject_FBP_multiGPU_splitViews, which creates per-chunk parameter copies via removeProjections(). Two bugs in this function caused incorrect reconstruction values:

Bug 1 (Major): phis_full stored in degrees, consumed as radians

removeProjections() calls get_angles(phis_full) to save the full-scan angle array. get_angles() converts from internal radians to degrees:
phis_full[i] = (phis[i] + 0.5*PI) * 180.0 / PI

Later, T_phi() preferentially reads phis_full over phis:
if (phis_full != NULL) phis_local = phis_full; // ← uses degrees!

Since T_phi() returns angular step without deg→rad conversion, and FBPscalar ∝ T_phi(), the normalization applied during Hilbert filtering is scaled by 180/π ≈ 57.3×. The derivative kernel (parallelRay_derivative) uses phis[] (always radians) for its local T_phi, so the 1/T_phi and ×T_phi do NOT cancel — the error propagates directly to reconstructed CT values.

Bug 2 (Minor): phi_start/phi_end restored to full-scan range

After set_angles(), phi_start/phi_end correctly reflect the chunk's angular range, but are immediately restored to full-scan values:
phi_start = phi_start_save; // full-scan
phi_end = phi_end_save;

These are copied to GPU constant memory (d_phi_start, d_phi_end) and used in the helical weighted backprojector kernel to bound the redundancy sumWeights computation. With full-scan bounds, sumWeights includes turns not present in the chunk's data, causing incorrect per-view normalization.

Repro Conditions

• Helical cone-beam scan (pitch ≠ 0)
• Projection data + volume data exceed available_GPU_memory (triggers backproject_FBP_multiGPU_splitViews)
• Example: 6000 views × 505 rows × 2063 cols = 23.3 GiB, vol 1527³ × 445 = 3.9 GiB, required ≈ 50.7 GiB > typical 48 GB GPU → triggered

Small datasets that fit entirely on GPU are NOT affected (no chunking, phis_full remains NULL, T_phi() uses phis[] in radians).

Verification

FBPscalar can be printed before FBP (via get_FBPscalar()) to verify:
FBPscalar = 1/(2π) × T_φ × pixW × (sod/sdd)² × pixH / (voxW² × voxH)

With the fix, FBPscalar ≈ 0.03–0.04 for typical micro-CT geometries (corresponding to T_φ ≈ 0.004 rad). Before the fix, chunked FBP would report the same FBPscalar value (computed from the pre-chunk params), but internally use a 57× larger value for each chunk.

Files Changed

• src/parameters.cpp: removeProjections() — store phis_full in radians, manually set phis and phi_start/phi_end without degree-expecting set_angles().