Why Lung Labeling Is Actually Worth Your Time
Most people coming into medical image annotation think they already understand it. They've seen the dashboards. They've watched a few demo videos. Then they actually open a dataset and realize how much of it is just figuring out whether a blur in the periphery is a blood vessel or a ground-glass opacity, and doing it consistently across 4,000 images without your brain starting to merge two different patients' scans together. I spent about eight months working on a lung CT labeling project for a radiology AI startup. We were building a nodule detection model, and the labeling guidelines looked clean on paper. They were a disaster in practice. Here's what actually happened and what I learned from it.
The Gamechanging World Of Labeling Lungs How A Simple Concept Is Revolutionizing Healthcare
At its core, lung labeling means taking medical images — chest X-rays, CT scans, sometimes MRIs — and attaching structured annotations to them. A bounding box around a nodule. A segmentation mask tracing the lung parenchyma. A classification tag like "consolidation present" or "pleural effusion." That's the simple part. The hard part is doing it well enough that a machine learning model can actually learn from it without absorbing all the noise and inconsistency that comes from human labelers. The reason this is becoming a bigger deal in healthcare is pretty straightforward. Every hospital system has thousands of unlabeled lung images sitting in PACS archives. Radiologists have been reading them for decades, but those reads exist as text reports, not as structured pixel-level data that an algorithm can consume. Labeling them properly unlocks that archive. It turns historical diagnostic work into training data. That's the revolution people are talking about, and it's not as abstract as it sounds. We used a tool called CVAT for the CT work and a custom labeling interface built on top of it for the X-ray batch. The workflow was: receive DICOM files, assign them to a labeler, the labeler marks regions of interest, a second annotator reviews and flags disagreements, and then a radiologist does final arbitration on the disputed cases. That pipeline took us about three weeks to stabilize. Before that, we were getting inter-annotator agreement rates in the low 60s on ground-glass opacities, which is useless for training a reliable model.
What Nobody Tells You About Lung Labeling
The first thing that hits you is that not all abnormalities look the same at different scan resolutions. We had a case where a nodule that was clearly visible on a high-resolution CT protocol looked like a vague density on a lower-resolution screen. One of our labelers flagged it as "uncertain" and moved on. The radiologist later confirmed it was a 4mm solid nodule. But because the labeler's confidence was low, it got tagged as a non-nodule in our training set. That single mislabel inflated our false-negative rate by about 3% on the test split. Small nodules under 6mm are basically invisible to most labeling workflows unless you're using thin-slice reconstructions and proper windowing settings. If you're not checking the original DICOM series and not testing different lung and soft tissue windows before labeling, you're going to miss things. A lot of things. The second thing is that labeler fatigue is real and it compounds fast. After about 90 minutes of continuous lung CT labeling, error rates climb noticeably. I tracked this ourselves by having a senior labeler do a second pass on a random 10% sample after a long session. The disagreement rate went from about 4% to 11%. We switched to 45-minute sessions with mandatory breaks and the consistency metric recovered. It's a small operational change but it saved us probably two weeks of rework.
Get the Full Details
A Practical Walkthrough
Start with the image source. If you're working with CT, make sure you have the original DICOM files, not converted JPGs or NIfTI files that have been resliced. DICOM preserves the HU (Hounsfield Unit) values, which matter enormously when you're trying to distinguish a fat-containing lesion from a fluid density. One of our early batches got fed converted images and we spent three days relabeling because the contrast windowing was baked in and wrong. Set up your labeling classes before you touch a single image. Don't wing it. For lung CT, a basic but workable schema includes: normal lung parenchyma, ground-glass opacity, consolidation, nodule (subdivide into solid and subsolid if possible), pleural effusion, pneumothorax, and emphysema. Each class needs a one-paragraph definition with an included example image. Labelers will interpret terms differently otherwise. "Ground-glass opacity" to a radiologist means something very specific. To a new labeler it might just mean "anything that looks hazy." Use polygon segmentation for diffuse processes like ground-glass opacities and emphysema. Use bounding boxes only for well-circumscribed nodules. Segmentation is slower but far more useful for training a model that needs to understand spatial extent. I've seen projects where they labeled everything with boxes because it was faster, then wondered why the model couldn't handle irregularly shaped consolidations.
Quality control should happen at three levels. First, automated checks — verify that labelers aren't tagging the entire lung field as a nodule, or creating masks that extend outside the body. Second, spot-checking — a senior labeler reviews 10-15% of new work daily. Third, radiologist arbitration — any case where the two annotators disagree, or where either one marks a confidence score below a set threshold, goes to a board-certified radiologist. This third step is where most projects either blow their budget or skimp too hard. Find a balance. For a typical project, having a radiologist review roughly 20% of cases caught about 85% of the labeling errors we encountered, and it cost a fraction of reviewing every single image.
Tools I'd Actually Recommend
For CT work, CVAT is solid and free. It handles DICOM import, supports polygon and box annotations, has a review workflow built in, and exports to COCO, YOLO, and segmentation mask formats. We ran it on a local server and it handled about 500 cases per day across three labelers without issues. For chest X-ray classification, LabelImg is fine for bounding box work but it's outdated for anything large-scale. I'd suggest Roboflow instead. It has better UI, handles class imbalance warnings, and lets you build a versioned dataset easily. The free tier covers small projects. If you need dedicated medical imaging tools, OHIF Viewer is open-source and works well with DICOM, though it requires more setup. We used it for the final radiologist review step because it loads series faster than CVAT and supports MPR (multi-planar reconstruction) natively, which matters when you need to scroll through axial, sagittal, and coronal views simultaneously to confirm a nodule isn't just overlapping rib structure.
For storing and managing the labeled datasets, DVC (Data Version Control) is the least painful option I've found. It tracks dataset versions alongside model training runs. Most teams I've worked with ignored data versioning until they needed it and spent two weeks scrambling to figure out which labeling iteration produced their best model. Don't be that team.
Where This Approach Breaks Down
Lung labeling doesn't work well when your imaging protocols are inconsistent across sites. We had a multi-center dataset where one hospital scanned at 5mm slice thickness and another at 1.25mm. The same pathology looked completely different between the two. Models trained on mixed-resolution data tend to overfit to the higher-resolution source and underperform on the coarser scans. The workaround was to train separate models per resolution and blend the outputs, but that adds complexity most teams aren't prepared for. Another failure mode is rare-pathology labeling. If your dataset has only 30 cases of a specific condition, no amount of labeling rigor will make a reliable model. You'll get a classifier that memorizes those 30 cases but fails on anything unseen. In those situations, the honest answer is to use data augmentation carefully, seek out collaborative datasets, or accept that the model will have limited utility for that specific diagnosis. There's no labeling trick that fixes a sample size problem. Cost is another real constraint. At the rates we were paying — roughly $2-4 per labeled CT case for standard annotation, $8-15 per case when radiologist arbitration was required — a project with 10,000 labeled cases was running about $60,000 to $100,000. That's before you account for the radiologist time, which is the expensive part. If you're working with a tight budget, prioritize quality over quantity. Five thousand well-labeled cases beat ten thousand rushed ones every time.
Bottom Line
Lung labeling is unglamorous work. It involves staring at thousands of cross-sectional images, making the same decision over and over, and dealing with the frustration of borderline cases that even experienced radiologists disagree on. But when done correctly, it produces something genuinely useful. Structured, high-quality labeled lung datasets are one of the most valuable inputs for medical AI right now, and the demand for them is only going to increase as more hospitals try to move from retrospective reads to prospective computational assistance. If you're planning to do this work, plan for more time and more money than you think you need. The edge cases will find you.
