Implement simpler API for optimizing intrinsic parameters

[eigenvivek]

The current focus of DiffDRR is on optimizing extrinsic parameters. This is perfectly fine for calibrated C-arms (e.g., research scanners), but it seems that clinical scanners are often uncalibrated! Image-based 2D/3D registration therefore requires optimization of the intrinsic parameters as well as the extrinsics. With the current DiffDRR API, this is cumbersome.

Proposed change:

• Instead of initializing the diffdrr.detector.Detector with a specific pixel spacing, principal point, and focal length, simply initialize with unit values (on the GPU). In the forward method, pass delx, dely, x0, y0, and sdd, and use these values to resize the unit detector. Perhaps this could be implemented as an affine matrix?
  

  DiffDRR/diffdrr/detector.py

  Lines 73 to 116 in bc2d869

  	def _initialize_carm(self: Detector):
  	"""Initialize the default position for the source and detector plane."""
  	try:
  	device = self.sdd.device
  	except AttributeError:
  	device = torch.device("cpu")
  	# Initialize the source at the origin and the center of the detector plane on the positive z-axis
  	source = torch.tensor([[0.0, 0.0, 0.0]], device=device)
  	center = torch.tensor([[0.0, 0.0, 1.0]], device=device) * self.sdd
  	# Use the standard basis for the detector plane
  	basis = torch.tensor([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0]], device=device)
  	# Construct the detector plane with different offsets for even or odd heights
  	h_off = 1.0 if self.height % 2 else 0.5
  	w_off = 1.0 if self.width % 2 else 0.5
  	# Construct equally spaced points along the basis vectors
  	t = (
  	torch.arange(-self.height // 2, self.height // 2, device=device) + h_off
  	) * self.delx
  	s = (
  	torch.arange(-self.width // 2, self.width // 2, device=device) + w_off
  	) * self.dely
  	if self.reverse_x_axis:
  	s = -s
  	coefs = torch.cartesian_prod(t, s).reshape(-1, 2)
  	target = torch.einsum("cd,nc->nd", basis, coefs)
  	target += center
  	# Batch source and target
  	source = source.unsqueeze(0)
  	target = target.unsqueeze(0)
  	# Apply principal point offset
  	target[..., 1] -= self.x0
  	target[..., 0] -= self.y0
  	if self.n_subsample is not None:
  	sample = torch.randperm(self.height * self.width)[: int(self.n_subsample)]
  	target = target[:, sample, :]
  	self.subsamples.append(sample.tolist())
  	return source, target