Marcus_Osei
Data Scientist · NHS Digital
Mar 2026
One important regulatory note for anyone deploying this in a clinical setting in the UK or EU — under the EU AI Act (which took full effect in August 2025), AI systems used for medical imaging segmentation in a clinical decision support role are classified as high-risk AI and require conformity assessment before deployment, separate from the CE marking under MDR/IVDR. The MONAI Deploy SDK helps with documentation requirements because it generates structured provenance logs for each inference run, which you need for the required technical documentation. The MHRA in the UK published updated AI as a Medical Device (AIaMD) guidance in Q1 2026 that's worth reading carefully — it clarifies when a radiology AI tool crosses from general purpose into regulated device territory. Don't assume that wrapping a model in a "decision support" label keeps you out of regulatory scope if it's influencing treatment pathways.

The sliding_window_inference with 50% overlap is important for small nodules near patch boundaries — I've seen people use 0.25 overlap and miss 6–8mm nodules consistently. The LUNA16 dataset is still the best public benchmark for validating your pipeline before clinical use.

Lena_Fischer
Medical Physicist · Charité Berlin
Feb 2026
Sanjay's point about spacing is critical and I want to add the code for it since getting this wrong is the single most common mistake I see in teams starting out. Here's a minimal working MONAI inference pipeline for a lung CT volume that handles spacing, windowing, and outputs a binary nodule mask:

Sanjay_Reddy ✓ Clinical AI
Radiologist + ML · AIIMS Delhi
Feb 2026
MONAI has become the de facto framework for medical imaging AI and the jump from v1.3 to v1.4 (released late 2025) was significant — the new Auto3DSeg pipeline in particular has made lung nodule segmentation accessible without needing deep ML expertise. The one thing nobody tells you upfront: always apply MONAI's built-in intensity windowing to lung window presets (-1000 to +400 HU) before feeding CT volumes into any segmentation model, because most pretrained checkpoints were trained with this normalization baked in and skipping it will tank your Dice scores by 15–20 percentage points. Also, make sure you use monai.transforms.Spacingd to resample everything to 1.5mm isotropic spacing — lung nodule models are extremely sensitive to voxel spacing mismatches. The official MONAI model zoo at monai.io/model-zoo has the lung nodule detection bundle with pretrained weights that work immediately on NIfTI formatted NLST-style data. For production deployment, the MONAI Deploy SDK wraps the inference pipeline into a containerized MAP (MONAI Application Package) that integrates cleanly with clinical PACS infrastructure.