lesiondiffusion-mlx 0.3.0

Breaking changes

run returns the shared medmlx-core result with a new artifacts entry, and the output roles are renamed. outputs keeps the same files under image (was synthetic_image), mask (was target_mask) and raw_image (was raw_synthetic_image); organ_mask is unchanged. artifacts describes each output, and details keeps the run metadata.

-from lesiondiffusion_mlx.medmlx import run
+from lesiondiffusion_mlx import run
 result = run(image_path="ct.nii.gz", organ_mask_path="liver.nii.gz", output_dir="out", organ="liver", lesion_type="tumor")
-print(result["outputs"]["synthetic_image"], result["outputs"]["target_mask"])
+print(result["outputs"]["image"], result["outputs"]["mask"])

Requires medmlx-core>=0.3,<0.4 (was 0.2.x).

Fixes

Outputs are restored if publishing the new files fails part-way.

New

load_model() loads the weights once for repeated draws on one CT and lesion outline, for example many seeds or phenotype texts. Each model.run(...) takes the same arguments as run and returns the same result; the validated geometry and encoded CT latent are reused.

sampler="dpmpp_2m" selects deterministic DPM-Solver++ 2M and lmnet_diffusion_steps / linet_diffusion_steps set the step count of each stage. The default 200-step DDIM path is unchanged, and source mode keeps DDIM.

decoder_mode="scaled_fp16" decodes in scaled half precision on M5-class GPUs with NAX enabled. Lesion-region HU differs from the default FP32 decoder by about 0.12 HU on average (maximum below 0.5 HU).

The README example now uses output_region="lesion", which keeps the input CT outside the lesion neighborhood. The default stays "full".

Faster

With sampler="dpmpp_2m" at 10 steps, a generated-mask run takes 3.5 s warm at 2.60 GB peak MLX memory with decoder_mode="scaled_fp16", against 4.7 s at 5.19 GB with the default FP32 decoder (M5 Max, 160x160x128 crop). The VQ encoder, LINet and VQ decoder now load one after another, which lowers the scaled path's peak.