Medical imaging interview questions is a Q&A list for an imaging-ML / imaging-engineer interview: files, cameras, metrics, models, and what you are not allowed to invent on a patient. It is not a bootcamp. It is not “must-know 2025.” It is not a career on-ramp.
If you meant which hole to fill first (CV vs clinician) → medical imaging starter. If you meant body → pixels, the four cameras → acquisition of images. If you meant imaging vs therapy, DICOM modality codes → modalities definition medical. If you meant the DICOM file → what is DICOM.
PYCAD builds custom web DICOM viewers and imaging models. It does not sell a course, a certificate, or an interview-prep product. The old MONAIGPT demo linked from this page is leftover marketing; it is not an answer key.
Files and cameras
What is medical imaging, in one sentence a hiring manager will accept?
A camera (or a tracer) writes a file of the inside of a body so someone can decide. Diagnosis, planning, follow-up, screening, and guidance are uses. “A picture of the inside” is the junior version; the file plus the protocol is the job.
Name the common techniques and what each is for.
X-ray / CR / DX: dense structures, chest, bone. CT: attenuation volume, trauma, lung, contrast vessels. MRI: water, soft tissue, no ionizing dose. Ultrasound: live echo, fetus, heart, a needle. Nuclear medicine (PET / SPECT): a tracer’s function, almost always fused. Mammography: a dedicated breast X-ray. Fluoroscopy: a live X-ray during a procedure. The clinical question picks the camera. Physics lives on acquisition.
What is DICOM and why does the interview keep asking?
DICOM is the file and the handshake: pixels plus tags (who, which series, which window, which modality code). A PNG of a slice is not a study. Lose the tags and you cannot hang, bill, or train honestly. Longer version: what is DICOM.
2D vs 3D vs 4D?
2D is one plane (a radiograph, one US frame). 3D is a volume (CT / MRI stack). 4D is that volume over time (cine cardiac, 4D CT for a moving target). Do not call a multi-slice CT “4D” because it has many slices.
What is multi-modal imaging?
Two (or more) cameras on the same question: PET/CT, PET/MR, T1+T2+FLAIR, a pathology slide next to the radiology series. Benefit is complementary contrast. Cost is registration, different grids, and a model that will cheat on the easier modality if you are sloppy.
Data
Missing or corrupted files?
Say what is missing first: a tag, a slice, a whole series, burned-in PHI, a failed reconstruct. Do not mean-impute a CT as if it were a spreadsheet. Drop the series, re-pull from PACS, or exclude the patient and write it down. Interpolation across slices invents anatomy. Augmentation is not a repair for a broken acquisition.
Class imbalance?
Rare positives are the default. Weighted loss, focal loss, oversample the rare class, undersample the easy negatives, or a two-stage detect-then-classify. Do not invent prevalence with a GAN and then report that AUC as clinical. Site shift usually hurts more than class ratio.
Preprocessing you should be able to name.
Orientation to a RAS/LPS convention. Spacing / resample (or refuse to). Intensity: HU is already a unit on CT; MRI is not — z-score or percentile inside a mask, not a global 0–1 on a volume with air. Window/level for display is not the same as the array you train on. Crop to a FOV you can defend. De-identify tags and burned-in pixels. Crop and preprocessing are different leftover/cluster URLs; do not recap them here.
Data augmentation — when is it honest?
Flips and small rotations if the anatomy still makes sense (a liver can swap sides in the array; a situs-inversus label cannot). Intensity jitter on MRI. Elastic deform only if you still trust the contour. Do not flip a chest X-ray and keep “left pleural effusion.” Do not claim a GAN sample is a held-out site.
Normalization vs standardization?
In this interview, “normalize” usually means put intensities on a comparable scale. CT: leave HU, or clip to a window and scale that window. MRI: per-volume or per-mask z-score. Do not min–max a whole study including the air and the coil-hot edge and call it done.
Large datasets / noisy images?
Stream, do not RAM-load. Cache decoded arrays. One .dcm is not a series. For noise: denoise only if you measured that it helps the task; a pretty slice can wipe a 3 mm nodule. Protocol and motion are decided at acquire — acquisition — not in a prompt.
Models
Why CNNs on images?
Local filters, weight sharing, a hierarchy from edges to organs. That is still the default inductive bias for a slice. Architecture explainer: CNNs explained. A real 2023 CNN-vs-ViT run (PNG X-rays, caveated): CNNs or ViT.
Transfer learning — what actually transfers?
A backbone trained on a large set (often ImageNet RGB) plus a new head. Freeze early layers when the set is small; unfreeze when you have enough and a low LR. Medical volumes are not ImageNet. A frozen RGB encoder on a single CT window is a starting point, not a paper. How-to with numbers: frozen-weight transfer.
Supervised / unsupervised / semi-supervised?
Supervised: every sample has the label you will be scored on (mask, box, class). Unsupervised: structure without that label (clustering, reconstruction). Semi-supervised: a few labels, a lot of unlabeled — consistency, pseudo-labels, or a pretext task. In clinic the label is the expensive part; say so.
What is image segmentation, and U-Net?
A label per pixel (or voxel): organ, tumor, vessel. U-Net (Ronneberger et al., 2015) is the encoder–decoder with skip connections that still ships for this. Semantic = class per pixel. Instance = this tumor vs that tumor. Do not quote an unsourced Dice as a law. Methods live in the 675 cluster, not this list.
Classification vs detection vs segmentation?
Classification: one label for the study or the crop (“pneumothorax: yes”). Detection: a box or a point (“where”). Segmentation: the outline. CAD is the clinical wrapper: a mark the reader can accept or reject — AI for medical diagnosis.
GANs in imaging?
A generator fakes, a discriminator calls. Uses: extra samples, domain translation, a reconstruction prior. They are not a site hold-out. 2D catalogue how-to: medigan. Text-to-3D implicit functions are a different URL: Shap-E.
Overfitting — what do you actually do?
A hold-out that is not a random slice of the same patient. Regularize, early-stop on a patient-level val, augment, simplify the net. If val Dice is 0.94 and an outside site is 0.61, you overfit the hospital, not the epoch count.
Radiomics vs a net?
Radiomics: a hand-specified feature table from a contour (shape, texture, histograms). A net learns the features. Both still need a reproducible contour and a site hold-out. Field piece: future of medical imaging.
Multi-task? RNNs? Reinforcement?
Multi-task: one backbone, two heads (segment + classify) when the labels share anatomy. RNNs / LSTMs: sequences — cine, a report, a time series — not a default for a static CT. Reinforcement: a policy (a probe path, a hanging-protocol choice). Rare in a first-round imaging screen; say so instead of a TED sentence.
Feature extraction?
Either a named radiomic / classical descriptor, or the latent of a trained net. “We extracted features” is not an answer. Name the layer or the table.
Metrics and mistakes
Precision, recall, F1 — in this job?
Precision: of the marks you fired, how many were real (PPV in a given prevalence). Recall / sensitivity: of the real ones, how many you caught. F1: the harmonic mean when you need one number. In screening, a missed cancer (FN) and a useless workup (FP) are not symmetric. Also name specificity, AUC, and — for masks — Dice / IoU and a Hausdorff, not only pixel accuracy.
How do you evaluate a model?
Split by patient / site / scanner, not by slice. Report the metric that matches the decision. Show the confusion and a few failure cases. Calibration if a probability is shown to a clinician. A leaderboard number without a protocol is theater.
False positives and false negatives?
Pick the operating point on a val set that is not the test set. A second reader or a rule (size, HU, clinical indication) is how you live with the leftover errors. Do not “fix” FN by flooding the worklist with FP.
Practice
Ethics and privacy?
Consent and the legal basis for the file. De-identify tags and pixels. Do not train on a production PACS export you were handed on a USB. Do not ship a model that leaks the training set. HIPAA / transfer is a different URL: HIPAA-compliant data transfer. Federated learning (the model travels) is this post, not a privacy fortress by itself.
Why sit with a clinician?
They set the question, the label, and whether a 0.02 Dice lift is a wasted year. You set the split, the leak, and the file. An interview that cannot name both sides is a red flag.
What this page is not
- Not 515. The on-ramp (which hole to fill) stays on that URL.
- Not a PYCAD course, bootcamp, or certificate. Those landings 404 or were never a product.
- Not leftover 528 / MONAIGPT. The Streamlit demo was an ad at the bottom of this post. Dropped.
- Not “40 must-know questions for 2025.” The list is the interview topics. Years are not a syllabus.
If the job later is a viewer or a model, that is the imaging piece. Case studies.