Locate the target, sample multiple positions, and combine CT features to predict AEA presence/absence and Keros type.
Check each scan, standardize it, and find the target region.
Load each DICOM series as a 3D CT volume.
Check orientation, spacing, slice thickness, and coverage. Create coronal, axial, and sagittal views.
Align orientation, spacing, and intensity settings. Preserve thin bone details.
Use anatomical landmarks to select relevant slices and crops.
Understand the anatomy before choosing the model input.
A CT volume is a 3D array. Each voxel is a small tissue volume with an intensity value, typically in HU.
D = slices; H / W = pixels along the other two axes.
200 × 512 × 512 ≈ 52 million voxels
Array-axis order depends on the loading library and orientation convention.
Hounsfield Units (HU) measure X-ray attenuation and help distinguish tissues.
Approximate ranges, not fixed tissue boundaries.
Spacing gives the distance between neighboring voxel centers along each axis.
Example spacing: (0.5, 0.5, 1.0) mm
Along the 1.0 mm axis, 5 voxel intervals = 5 mm.
DICOM (.dcm): clinical images plus orientation, position, pixel spacing, slice/reconstruction details, and HU conversion parameters.
NIfTI (.nii / .nii.gz): a convenient 3D volume with spatial metadata for machine learning.
Preserve physical geometry and verify HU conversion when loading or converting files.
Align anatomical axes.
Use consistent spacing; preserve geometry.
Keep target anatomy; reduce background.
Emphasize relevant tissue, such as bone.
Scale intensities for the CNN.
Nasal cavity, ethmoid sinuses, and nearby structures.
Ethmoid roof, cribriform plate, and lateral lamella.
Anterior–posterior relationships.
Keros and AEA targets occupy small regions. Random slices or uniform whole-volume sampling may miss them.
The AEA’s course relative to the skull base matters in sinus surgery. A suspended course may increase exposure to vascular injury.
Runs close to or along the skull base.
Suspended or “floating” below the skull base.
Predict presence (1) or absence (0) using the project’s AEA labels. The two courses above provide anatomical context.
Use coronal CT with bone-window settings to locate the AEA canal and assess its relationship to the ethmoid roof. Check adjacent slices and other planes to follow its course.
Keros describes olfactory fossa depth using lateral lamella height. This thin bone connects the cribriform plate to the ethmoid roof. Type III has the deepest fossa and greater risk of skull-base injury during sinus surgery.
| Keros type | Height |
|---|---|
| Type I | 1–3 mm |
| Type II | 4–7 mm |
| Type III | 8–16 mm |
AEA and the lateral lamella occupy tiny regions in a full CT.
Implication: use anatomical localization to reduce irrelevant input.
Coarse segmentation may not locate the exact target.
Implication: test fine localization with physician annotations and transfer learning.
One 2D slice may not describe a small 3D structure.
Implication: investigate three views, adjacent slices, and local 3D features.
A few sample positions may still miss decisive anatomy.
Implication: test sampling density and aggregation. More samples may improve coverage but also add noise.
~33 patient CTs have clinical labels; detailed anatomical annotations may be missing.
Implication: prioritize pretrained features, lightweight classifiers, and patient-level cross-validation.
Locate → sample → extract features → combine → classify.
Validate location, then test sampling and feature aggregation.
Test TotalSegmentator on a few scans. Crop with a margin in mm and check target coverage.
For Keros, create separate L / R crops.
Use physician annotations to adapt a pretrained model: boxes → detector; points → landmark model; masks → segmenter.
Output: a validated fine 3D ROI.
Densely sample the fine ROI. At each position, extract aligned axial, coronal, and sagittal views, or nearby slice stacks.
Start with a shared, frozen ResNet18. Check coverage of physician-marked critical regions.
Fuse views at each position, then combine positions with Mean or Top-K pooling. Compare with Attention MIL.
Train small classifiers for AEA presence/absence and Keros I/II/III per side.
Each task uses its own target region and output labels.
Binary output: presence (1) or absence (0), using the dataset’s AEA labels.
Main approach: predict Type I, II, or III separately for L and R. Backup: measure height, then assign the type.
Check target location first, then compare classifiers.
Compare predicted regions with physician annotations.
MedSAM: test whether physician prompts give useful regions.
Boxes / points: detector or landmark model.
Masks: nnU-Net or a MONAI segmentation model.
B0: Coronal view + frozen ResNet18 + linear classifier.
B1: One tri-planar sample + shared ResNet18.
B2: Dense sampling + Mean / Top-K / Attention MIL. Compare with B1 to test coverage gains.
Use patient-level cross-validation. Keep each patient’s slices and sides together. Learn preprocessing settings and tune models only on training folds.
Find physician-defined landmarks in the coronal view. Use CT spacing and orientation to calculate height.
Steps: target crop → landmarks → height (mm) → Keros type.
Try this if direct classification is unreliable or the physician prefers measurements.
Concern: Can TotalSegmentator find the regions needed for Keros and AEA?
Validation: Test several CTs. Visually check whether the generated ROIs fully contain the targets.
Concern: Can limited physician labels locate very small structures within the coarse ROI?
Validation: Compare predicted boxes with physician ground truth: localization error (mm), IoU, and target coverage.
Concern: Do three views capture enough fine anatomical detail?
Alternative: Test multi-slice 2.5D CNNs, lightweight 3D CNNs, or feature-level multi-view fusion.
Validation: Compare coronal-only, tri-planar, and models with more spatial context.
Concern: One tri-planar sample may miss the anatomy that determines the label.
Approach: Densely sample candidate positions in the fine ROI. Extract aligned axial, coronal, and sagittal views at each position with a shared pretrained CNN.
Aggregation: Combine position features with Mean Pooling, Top-K Pooling, or Attention-based Multiple Instance Learning (MIL).
Validation: Check coverage of physician-marked critical regions. Compare aggregation methods and test whether irrelevant samples affect predictions.
Concern: Direct CNN classification may be less reliable than measuring lateral lamella height, which defines Keros type.
Alternative: If classification performs poorly, detect landmarks and measure height in mm, then assign Keros type for L and R.