Why a computer-vision expert’s intuitions misfire

If you have trained a model on ImageNet, COCO, or a few hundred million Instagram photos, you have excellent instincts for natural-image vision. Most of those instincts are wrong — or at least dangerously incomplete — the moment you point them at a chest CT or a screening mammogram.

This post is a field guide for machine-learning scientists moving into radiology. It is not a survey of architectures; the architectures are mostly the ones you already know (CNNs, U-Nets, vision transformers, increasingly foundation models). What changes is everything around the architecture: the statistics of the signal, the cost and meaning of a label, the data you can actually get, and — the part that quietly sinks most projects — generalization across the bewildering heterogeneity of how medical images are produced. I will end with the two things ML scientists most often discover too late: how the FDA actually regulates these models, and why the model in the paper is rarely the model that ships.

A running theme: medical imaging is in some ways easier than natural-image vision, and leaning on those advantages is the difference between a model that demos well and one that survives contact with a second hospital.

What is genuinely easier than natural images

Start with the good news, because it is real and underexploited.

Canonical pose and framing. A street scene can contain a cat at any scale, any orientation, anywhere in the frame, against any background. A PA chest radiograph is, by protocol, a patient standing upright, facing the detector, arms positioned to rotate the scapulae off the lung fields. The heart is on the left.¹ The aortic knob is where the aortic knob goes. This is a strong spatial prior that natural-image models simply do not get for free — and it is why registration, atlas-based priors, and even fixed positional encodings work far better here than they would on web images.

One channel, calibrated. Most modalities are grayscale, and — crucially — the gray values often mean something physical. CT is quantitative: each voxel is a Hounsfield unit, a linear transform of the X-ray attenuation coefficient \(\mu\) relative to water,

so water is \(0\), air is \(-1000\), fat is around \(-100\), and cortical bone is \(+1000\) or more. Fat is fat in every CT scanner on Earth. Nothing in RGB is calibrated like this; “how blue is the sky” is not a physical constant. You can and should exploit it — windowing, HU-based preprocessing, and physically motivated augmentations all follow from it.

The suspected disease localizes attention. Clinical imaging arrives with a reason for exam. “Rule out pneumothorax” tells you to look at the pleural line; “rule out stroke” sends you to the brain parenchyma and vessels. The organ of interest is usually known, which is a luxury object detection never has.

But each of these advantages has a barb:

• The canonical pose breaks for portable/supine films, pediatric patients, body habitus, and post-surgical anatomy.
• HU calibration drifts with scanner, kVp, and contrast timing (more on this below), and MRI intensities are not standardized at all — a T1 value is only meaningful relative to the rest of that one acquisition.
• “The organ of interest is known” is a trap: incidental findings in the other organs are often what matter most clinically. The lung-nodule model that ignores the adrenal mass at the edge of the field has failed the patient even if its AUC is perfect.

So: use the priors, but treat every one of them as a covariate that can shift.

The needle in the haystack: subtlety and extreme imbalance

Here is the single biggest statistical difference from natural images. In COCO, the object you care about typically occupies a meaningful fraction of the frame. In radiology, the finding is often a handful of voxels in a sea of normal tissue, and the difference between malignant and benign — between call the patient back and see you in two years — can come down to a few millimeters of spiculation or a subtle change in density.

Make it concrete with geometry. A chest CT of roughly \(512 \times 512 \times 320\) voxels at \(0.7 \times 0.7 \times 1.0\,\text{mm}\) contains about \(8.4 \times 10^7\) voxels. A clinically important \(5\,\text{mm}\) pulmonary nodule is a sphere of volume \(\tfrac{4}{3}\pi r^3 \approx 65\,\text{mm}^3\), or about \(134\) voxels. The lesion is therefore

of the volume — roughly one in six hundred thousand voxels. Shrink it to a \(3\,\text{mm}\) nodule and you are at one in three million. Figure 1 puts several findings on the same axis as natural-image objects; note the five-to-six order-of-magnitude gap.

The consequences for an ML scientist are direct:

• Accuracy is meaningless and pixel-wise loss is treacherous. A segmentation model that predicts “no lesion” everywhere achieves \(1 - 1.6\times10^{-6} \approx 99.9998\%\) voxel accuracy. Use overlap and detection metrics built for imbalance — Dice / \(F_1\), where for prediction \(P\) and ground truth \(G\), \(\mathrm{Dice} = \frac{2|P \cap G|}{|P| + |G|},\) free-response ROC (FROC) for detection, and class-balanced or region-based losses (Dice loss, Tversky, focal). The focal loss down-weights the easy negatives that otherwise dominate the gradient: \(\mathrm{FL}(p_t) = -(1-p_t)^{\gamma}\log p_t\).
• Most of the volume is uninteresting, and uninteresting in a structured way. Hard-negative mining, lesion-aware patch sampling, and two-stage candidate-then-classify pipelines exist because uniformly sampling voxels wastes almost all of your compute on obvious lung parenchyma.
• Resolution is not negotiable. Downsampling a natural image to \(224^2\) loses a cat’s whiskers; downsampling a CT slice can erase the lesion entirely. The signal you are hunting may be at the Nyquist limit of the acquisition.

Annotation is the bottleneck, not the model

In natural-image land, labels are cheap: crowdworkers draw boxes, and “is this a dog” needs no credential. Radiology inverts this completely, and it reshapes what is feasible.

A bounding box is the wrong primitive, and often impossible. Many findings have no crisp boundary. Where exactly does a ground-glass opacity end and normal lung begin? What is the bounding box of diffuse interstitial disease, or of “the lungs look hyperinflated”? The pathology is frequently a texture or a global property, not a localizable object. Even when a lesion is discrete, it lives in 3D — a box becomes a volume, and a radiologist scrolling 320 slices to contour a tumour is spending clinical time that costs orders of magnitude more than a crowdworker.

Ground truth is noisy and sometimes unobtainable from the image alone. The honest label often is not in the pixels. Is that lung nodule malignant? The image cannot say; you need the biopsy, or two years of follow-up showing growth. This is why so many “labels” in public datasets are actually NLP-extracted from the radiology report (MIMIC-CXR, CheXpert, ChestX-ray14, PadChest all do this) — which means your labels inherit both the radiologist’s error rate and the text-mining model’s error rate.

Inter-reader variability is a hard ceiling. Radiologists disagree. The LIDC-IDRI lung-nodule database was annotated by four thoracic radiologists precisely because no single read is ground truth; of 2,669 lesions marked as nodules \(\geq 3\,\text{mm}\) by at least one reader, only about 35% were marked by all four. If your “ground truth” is one radiologist, your evaluation noise floor may be larger than the improvement you are claiming. Model the labels as noisy: capture annotator agreement (e.g. Cohen’s / Fleiss’ \(\kappa\)), train against multi-reader consensus where you can, and report performance relative to the inter-reader band, not to an imagined perfect oracle.

You cannot read the data without domain knowledge. A computer-vision engineer can sanity-check an ImageNet pipeline by eye. Almost no ML scientist can look at a FLAIR hyperintensity and tell whether the label is right. This has a practical implication that teams underestimate: you need a radiologist in the loop continuously, not just at the start, because data-cleaning decisions (which views to keep, how to handle priors, what counts as positive) are clinical judgments in disguise.

The data scarcity problem

Natural-image research rides on ImageNet (\(1.4\)M images), and webscale sets in the billions. Radiology has nothing remotely comparable that is public, and the reasons are structural: images are protected health information, they must be de-identified (including burned-in pixel annotations and faces reconstructable from head CT/MRI), and the expert labels are expensive. What we do have is a handful of landmark public collections, summarized in Table 1.

Table: Major public medical-imaging datasets. “Images” counts vary by modality (a CT/MRI “study” is a 3D volume of many slices). Sizes are as reported by the source publications.

Two things to internalize. First, the largest labeled sets are 2D chest radiographs, because they are the cheapest to acquire and the easiest to label from reports; 3D, multi-sequence, and rarer-modality data are one to three orders of magnitude smaller. Second — and this is the setup for the rest of the post — a big total \(N\) is not the same as a big \(N\) where it counts. EMBED has 3.4M images, but if you want to evaluate performance for, say, architectural distortion in dense breasts of women under 40 scanned on one vendor’s tomosynthesis unit, you are suddenly working with a few dozen cases.

Heterogeneity and generalization: the part everyone underestimates

Everyone says medical-imaging AI “doesn’t generalize.” Fewer people say why, mechanistically. The reason is that a medical image is the output of a long physical and human pipeline, and every stage of that pipeline is a covariate that differs across hospitals. A natural image has confounders too (lighting, camera), but nothing like this stack.

Formally, the trouble is distribution shift. Your model learns \(P_{\text{train}}(Y \mid X)\) over inputs drawn from \(P_{\text{train}}(X)\), and is deployed where both can differ:

Decompose it. Covariate shift is \(P(X)\) changing while \(P(Y\mid X)\) holds — a different scanner renders the same pathology with different texture. Label shift is \(P(Y)\) changing — disease prevalence differs across a referral center and a screening clinic, which (via Bayes) moves every predicted probability and every PPV even if the imaging is identical. Concept shift is the genuinely dangerous one, \(P(Y\mid X)\) itself changing — the imaging appearance of a disease differs by population, or the label definition differs by institution. Here is the catalogue of what actually shifts:

• Scanner vendor and model. GE, Siemens, Philips, Canon detectors and reconstruction software impose vendor-specific texture and noise signatures. Models readily learn the scanner, not the disease.
• Acquisition physics. CT: tube voltage (kVp), tube current (mAs), pitch, slice thickness, and especially the reconstruction kernel (sharp vs. smooth) dramatically change texture — reconstruction kernel alone can render the majority of radiomic features non-reproducible across settings. MRI: field strength (1.5T vs 3T), pulse sequence and vendor implementation, TR/TE, and the fact that intensities are not standardized at all.
• Contrast and timing. With vs. without IV contrast, and when in the contrast bolus the scan was captured, can change a structure’s appearance more than disease does.
• Imaging noise and dose. Low-dose protocols (and the shift toward them) raise quantum noise; denoising and dose vary by site and by patient size.
• Patient demographics and disease spectrum. Age, sex, body habitus, ancestry, comorbidity mix, and disease prevalence and severity all vary by catchment. A model tuned where pneumothoraces are large and obvious degrades where they are small and subtle.
• Protocol and positioning. Portable vs. fixed units, supine vs. upright, inspiration depth, pediatric protocols, post-surgical hardware.

The canonical demonstration is Zech et al. (2018): CNNs trained to detect pneumonia on chest radiographs generalized worse to outside hospitals than internal test performance suggested, and the models had learned to detect the hospital system and even the department — exploiting that a portable scanner marker or a prevalence difference correlated with disease. The same pattern shows up in segmentation: AlBadawy et al. (2018) found glioma-segmentation performance dropped measurably when training and test institutions differed. This is shortcut learning, and it is rampant precisely because the spurious features (scanner, view, burned-in markers) are so predictable.

What this means for your workflow:

• Internal test performance is an upper bound, not an estimate. The only trustworthy evaluation is external — a held-out site, ideally a held-out vendor and time period. Split by hospital, not by image.
• Audit for shortcuts. Saliency maps that point at the corner marker, an AUC that survives when you black out the anatomy, a model that can classify scanner from the image — all are red flags.
• Harmonize deliberately. Intensity normalization, resampling to common spacing, vendor-aware augmentation, and even learned kernel/stain-style conversion exist to fight covariate shift; use them, but verify they did not erase the signal.

The statistical-power trap, in numbers

Now combine the previous two sections — heterogeneity and scarcity — and you get the quietest failure mode in the field. To prove a model generalizes, you must evaluate it in each clinically relevant subgroup. But every stratification you add slices your sample, and because disease is rare, it is the positive cases that vanish first.

Walk it down for a chest-radiograph model, anchored to MIMIC-CXR’s 377,110 images (Figure 2). Keep frontal views only (\(\times 0.65\)). Keep the positives for your target finding — pneumothorax, prevalence \(\approx 3\%\) (\(\times 0.03\)); already you are at ~7,000 positive cases, not 377,110. Now ask the generalization questions clinicians will ask: how does it do in women (\(\times 0.47\)), specifically those aged 18–40 (\(\times 0.16\)), specifically scanned on vendor B (\(\times 0.30\)), specifically with the moderate-to-large, actionable subtype (\(\times 0.40\))? You land on about 66 positive cases. From 377,110 to 66 — and 66 is the number that actually governs what you can conclude about that subgroup.

Why 66 is a problem is pure sampling theory. Estimate a subgroup sensitivity (true positive rate) \(\hat{p}\) from \(n\) positive cases; its standard error is \(\sqrt{p(1-p)/n}\), so the 95% confidence half-width is about

At a true sensitivity of \(0.85\) and \(n = 66\), that half-width is \(\pm 0.086\): your estimate is “somewhere between \(0.76\) and \(0.94\).” You cannot distinguish a clinically excellent \(0.90\) from a borderline \(0.78\). (For small \(n\) use the Wilson interval rather than this normal approximation — the qualitative story is the same, and at these counts it matters.) Figure 3a shows the half-width shrinking only as \(1/\sqrt{n}\); the subgroup strata are marked.

Worse, suppose you want to detect a real subgroup gap — say sensitivity drops from \(0.85\) overall to \(0.75\) in young women on vendor B. The number of positives per group needed for a two-sided test at \(\alpha = 0.05\) with power \(1-\beta\) is

which for \(p_1=0.85,\, p_2=0.75\) works out to about 250 positive cases per group for 80% power. Your subgroup has 66, which buys roughly 30% power (Figure 3b): a two-in-three chance of missing a real, clinically meaningful degradation. And if you honestly test across, say, ten subgroups, a Bonferroni correction to \(\alpha = 0.005\) pushes the requirement to ~425 per group — while simultaneously, not correcting means some of your “significant” subgroup findings are noise. You are squeezed from both sides.

The lesson is not “give up.” It is to plan evaluation as a power calculation from day one: decide which subgroups are non-negotiable, estimate the positive counts you will actually have, and either acquire enough cases (often via multi-site collaboration) or state honestly which subgroups you are not powered to certify. Silent truncation — reporting one headline AUC computed over a population you never stratified — is how models that look published-ready fail in deployment.

How these models are actually regulated

If your model will touch patient care in the US, it is almost certainly a medical device, and the FDA’s framework shapes your engineering. A few facts ML scientists are routinely surprised by:

• Radiology dominates. From the 1990s through the mid-2020s, roughly three-quarters of all FDA-authorized AI/ML-enabled devices are in radiology — by far the largest category. This is your field.
• Almost everything clears via 510(k), not clinical trials. The dominant path is the 510(k), which establishes “substantial equivalence” to a legally marketed predicate device — not a randomized trial. (Genuinely novel devices use the De Novo path; the highest-risk ones need full premarket approval, PMA, which is rare for imaging AI.) A consequence: fewer than a third of FDA-authorized radiology AI devices have published prospective clinical testing. Substantial equivalence is a regulatory claim, not evidence your model helps patients — keep those separate in your head.
• Models had to be “locked.” Historically the FDA cleared locked algorithms — same input, same output, no learning in the field — because a continuously adapting model breaks the entire premarket paradigm.

What changed recently is worth knowing, because it directly affects how you can plan model updates. In December 2024 the FDA finalized guidance on the Predetermined Change Control Plan (PCCP). The idea: in your original submission, you pre-specify what you will be allowed to change (e.g. retrain on new sites, recalibrate a threshold), the methodology you will use to develop and validate each change, and an impact assessment — and then you can ship those pre-authorized modifications without a new marketing submission. For an ML scientist this is the bridge from “frozen forever” toward “responsibly updatable,” and it explicitly asks you to think up front about intended-use populations (ethnicity, sex, disease severity) and deployment environments. In practice it means your monitoring and revalidation plan is part of the product, not an afterthought.

The academic model is not the deployed model

Finally, the gap that ends the most promising projects. The model in the paper and the model in the hospital are different artifacts, optimized against different objectives.

Concretely, what bites teams crossing this gap:

• Operating point, not the whole curve. A clinician runs your model at one threshold. A great ROC curve with no defensible, calibrated operating point is not deployable. And because prevalence differs by site (label shift), the threshold that gives the right PPV in your lab is wrong in the clinic; plan to recalibrate, e.g. with Platt scaling or isotonic regression, per site.
• The long tail is the job. Benchmarks delete the ambiguous and corrupted cases that dominate a real PACS queue. In deployment those are the workload: the lateral mistakenly sent as frontal, the patient with prior surgery, the motion-degraded study. Your model needs a calibrated “I don’t know.”
• Prospective \(\neq\) retrospective. Retrospective AUC routinely overstates prospective performance; the few prospective and randomized radiology-AI studies have repeatedly come in below their retrospective hype.
• Automation bias and workflow effects. A deployed model changes radiologist behavior — sometimes it catches misses, sometimes it anchors the reader to a wrong call. The endpoint that matters is reader + model, not the model in isolation.
• Drift and monitoring. Scanners get replaced, protocols change, populations shift. A model that was validated in 2024 is not automatically valid in 2027. The PCCP framework above exists precisely because this drift is inevitable.

Takeaways

If you remember five things moving from natural images to radiology:

1. Exploit the priors, distrust them. Canonical pose, calibrated intensities, and a known organ of interest are real gifts — but each is a covariate that shifts, and the finding may be in the organ you weren’t told to look at.
2. Your signal is a needle. Lesions are \(10^{-7}\)–\(10^{-5}\) of the image. Abandon accuracy and pixel-wise loss; use detection/overlap metrics, imbalance-aware losses, and lesion-aware sampling, and don’t downsample away the disease.
3. Labels are the bottleneck. They are expensive, noisy, NLP-mined, and bounded by inter-reader disagreement. Keep a radiologist in the loop and model the label noise explicitly.
4. Generalization is the whole game. Split by site/vendor/time, hunt for shortcuts, and treat internal test numbers as upper bounds.
5. Power your evaluation before you train. Stratification destroys positive counts; decide which subgroups you can certify, and say so honestly. Then remember the deployed model lives at one calibrated operating point, under FDA rules, drifting over time — design for that from the start.

See the accompanying notebook.ipynb for the geometry, the stratification waterfall, the power calculations behind Figures 1–3, and an automated check that every citation below resolves.

References

1. Clark K, Vendt B, Smith K, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digit Imaging. 2013;26(6):1045–1057. doi:10.1007/s10278-013-9622-7
2. Johnson AEW, Pollard TJ, Berkowitz SJ, et al. MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports. Sci Data. 2019;6:317. doi:10.1038/s41597-019-0322-0
3. Irvin J, Rajpurkar P, Ko M, et al. CheXpert: a large chest radiograph dataset with uncertainty labels and expert comparison. Proc AAAI. 2019;33(01):590–597. doi:10.1609/aaai.v33i01.3301590
4. Wang X, Peng Y, Lu L, et al. ChestX-ray8: hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. CVPR. 2017:3462–3471. doi:10.1109/CVPR.2017.369
5. Bustos A, Pertusa A, Salinas J-M, de la Iglesia-Vayá M. PadChest: a large chest x-ray image dataset with multi-label annotated reports. Med Image Anal. 2020;66:101797. doi:10.1016/j.media.2020.101797
6. Armato SG III, McLennan G, Bidaut L, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans. Med Phys. 2011;38(2):915–931. doi:10.1118/1.3528204
7. Bakas S, Akbari H, Sotiras A, et al. Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features. Sci Data. 2017;4:170117. doi:10.1038/sdata.2017.117
8. Menze BH, Jakab A, Bauer S, et al. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Trans Med Imaging. 2015;34(10):1993–2024. doi:10.1109/TMI.2014.2377694
9. Knoll F, Zbontar J, Sriram A, et al. fastMRI: a publicly available raw k-space and DICOM dataset of knee images for accelerated MR image reconstruction using machine learning. Radiol Artif Intell. 2020;2(1):e190007. doi:10.1148/ryai.2020190007
10. Jeong JJ, Vey BL, Bhimireddy A, et al. The EMory BrEast imaging Dataset (EMBED): a racially diverse, granular dataset of 3.4 million screening and diagnostic mammographic images. Radiol Artif Intell. 2023;5(1):e220047. doi:10.1148/ryai.220047
11. Littlejohns TJ, Holliday J, Gibson LM, et al. The UK Biobank imaging enhancement of 100,000 participants: rationale, data collection, management and future directions. Nat Commun. 2020;11:2624. doi:10.1038/s41467-020-15948-9
12. Zech JR, Badgeley MA, Liu M, et al. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLoS Med. 2018;15(11):e1002683. doi:10.1371/journal.pmed.1002683
13. AlBadawy EA, Saha A, Mazurowski MA. Deep learning for segmentation of brain tumors: impact of cross-institutional training and testing. Med Phys. 2018;45(3):1150–1158. doi:10.1002/mp.12752

Reproduce all analyses in this post here.