Tracing Like a Clinician: Anatomy-Guided Spatial Priors for Cephalometric Landmark Detection

Pallavi Mohanty; Sidhartha Mohapatra

arxiv: 2605.03358 · v3 · pith:YR5WIBWUnew · submitted 2026-05-05 · 💻 cs.CV

Tracing Like a Clinician: Anatomy-Guided Spatial Priors for Cephalometric Landmark Detection

Sidhartha Mohapatra , Pallavi Mohanty This is my paper

Pith reviewed 2026-07-01 00:38 UTC · model grok-4.3

classification 💻 cs.CV

keywords cephalometric landmark detectionspatial priorsanatomy-guided pipelineHRNetmedical image regularizationradiograph analysistraining-time prior

0 comments

The pith

Anatomy-guided spatial priors used only at training time cut cephalometric landmark error to 1.04 mm while closing the validation-to-test gap.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that a five-phase pipeline can turn a clinician's anatomical tracing workflow into confidence-weighted spatial priors that regularize HRNet-W32 training on cephalometric radiographs. These priors yield 1.04 mm mean radial error across 25 landmarks in 1,502 images from seven or more devices. A controlled training-by-inference matrix shows that only image-specific anatomically correct priors achieve the result and the small generalization gap; random or absent priors do not. The priors function solely as a training regularizer and require no generation step at deployment. Cross-validation, permutation tests, activation maps, and downstream clinical measurements all converge on the same performance lift, which also appears in echocardiography and spine radiographs when landmark distributions have high spatial entropy.

Core claim

Only image-specific anatomically correct priors produce the 1.04 mm result, functioning as a training-time regularizer requiring no automated prior generation at deployment. The training x inference prior matrix isolates this mechanism: anatomical priors maintain a 1% validation-to-test gap versus 88% without priors, despite identical validation convergence; the expanded architecture alone provides no benefit; random priors yield partial but unstable improvement.

What carries the argument

Five-phase anatomy-guided pipeline that produces confidence-weighted spatial priors to shape HRNet-W32 training.

If this is right

Anatomical priors reduce the validation-to-test performance gap from 88% to 1% while random priors give only partial and unstable gains.
The architecture expansion alone produces no accuracy benefit without the correct priors.
The same prior mechanism improves landmark detection in echocardiography, cervical spine, and hand radiographs when spatial entropy is high.
All trained models remain inference-independent once the priors have shaped the weights.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Training-time anatomical regularization may be useful for other structured medical landmark tasks where inference must stay lightweight.
The observed scaling of prior benefit with spatial entropy suggests a way to decide in advance whether a new imaging domain will respond to this approach.
Because the priors act only during training, the method could be retrofitted to existing networks without changing their deployment footprint.

Load-bearing premise

The five-phase pipeline generates priors that accurately reflect true anatomical structures in each radiograph without systematic bias or over-constraint.

What would settle it

Replacing the anatomy-guided priors with spatially plausible but anatomically incorrect priors and observing whether the 1.04 mm error and 1% gap both disappear would falsify the claim that anatomical correctness is required.

Figures

Figures reproduced from arXiv: 2605.03358 by Pallavi Mohanty, Sidhartha Mohapatra.

**Figure 1.** Figure 1: System architecture. Five phases progressively extract anatomical information. Phases A–C use learned segmentation; Phase D applies zero-parameter geometric rules from clinical definitions; Phase E generates confidence-weighted spatial priors. Total Stage 0: ∼40 ms on CPU view at source ↗

**Figure 2.** Figure 2: Phase A on four test cases spanning imaging de view at source ↗

**Figure 3.** Figure 3: Phase B: Five anatomical zones with region-specific contrast enhancement. From left: cranial base (aggressive view at source ↗

**Figure 4.** Figure 4: Phase D results on three datasets (ISBI, CEPHA29, DentalCepha). Left panels: simplified contours color-coded by view at source ↗

**Figure 4.** Figure 4: Phase D results on three datasets (ISBI, CEPHA29, DentalCepha). Left panels: simplified contours color-coded by [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Phase E: Confidence-weighted attention maps on two cases (ISBI left, CEPHA29 right). Top-left: 25 predicted land view at source ↗

**Figure 5.** Figure 5: Phase E: Confidence-weighted attention maps on two cases (ISBI left, CEPHA29 right). Top-left: 25 predicted land [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 6.** Figure 6: Three-way ablation. Left: test MRE across con view at source ↗

**Figure 6.** Figure 6: Three-way ablation. Left: test MRE across con [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

**Figure 7.** Figure 7: Predicted landmarks on four representative cephalograms spanning three imaging devices. Colors indicate confidence view at source ↗

**Figure 8.** Figure 8: Knowledge distillation failure: validation MRE de [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗

**Figure 8.** Figure 8: Cross-dataset generalization. Performance is consis view at source ↗

**Figure 7.** Figure 7: Per-landmark ensemble improvement (3-model av [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

**Figure 9.** Figure 9: Training×inference prior matrix. Left: all rows are uniform (inference-independent). Right: only anatomical priors during training produce the best result; the 28-channel architecture alone (Zero 25-ch) matches the 3-channel baseline. Population-mean priors are worse than no priors, confirming that static spatial bias hurts generalization [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

**Figure 10.** Figure 10: Grad-CAM activation comparison at fuse_conv.0 for three landmarks of increasing difficulty. Each row: original cephalogram, Grad-CAM with anatomical priors, Grad-CAM without priors, and Stage 0 attention channel. ANS (Easy): With priors, activation concentrates on the anterior nasal spine; without priors, activation bleeds to the image border. Go (Medium): With priors, tight focus on the mandibular angle;… view at source ↗

**Figure 11.** Figure 11: Predicted landmarks on four representative cephalograms spanning three imaging devices. Colors indicate confi [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗

**Figure 12.** Figure 12: Cross-dataset generalization. Quantitative perfor [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗

read the original abstract

Clinicians trace cephalometric radiographs following a structured anatomical workflow, yet no prior system encodes this into computation. We present a five-phase anatomy-guided pipeline producing confidence-weighted spatial priors that shape HRNet-W32 training, achieving 1.04 mm mean radial error on 25 landmarks across 1,502 radiographs from 7+ imaging devices. A training x inference prior matrix isolates the mechanism: anatomical priors maintain a 1% validation-to-test gap versus 88% without priors (1.94 mm), despite identical validation convergence. The matrix establishes that all trained models are inference-independent, the expanded architecture alone provides no benefit, random priors yield partial but unstable improvement (1.72 mm), and only image-specific anatomically correct priors produce the 1.04 mm result -- functioning as a training-time regularizer requiring no automated prior generation at deployment. Five-fold cross-validation (p=0.0015), patient-level permutation testing (p<0.0001, n=151), quantified Grad-CAM analysis (88% vs. 74% in-zone activation, p<0.001), and clinical measurement validation (skeletal classification kappa=0.79-0.84, zero Class II<->III reversals, ICC>0.95) provide converging evidence. Cross-domain experiments on echocardiography, cervical spine, and hand radiography support the hypothesis that prior effectiveness scales with the spatial entropy of the landmark distribution.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper's real contribution is the training x inference matrix showing that only image-specific anatomical priors (not random or architecture alone) drive the 1.04 mm error as a training regularizer, with solid multi-line evidence, though the five-phase prior pipeline still needs a direct independence check.

read the letter

The key point is that clinician-style anatomical priors applied only at training time get HRNet-W32 to 1.04 mm mean radial error on 25 landmarks across 1502 radiographs, while the matrix isolates that the gain requires anatomically correct, image-specific priors rather than any spatial signal.

What is new is the matrix design itself. It directly compares anatomical priors against random priors, no priors, and architecture-only conditions at both training and inference stages. This setup shows the performance lift depends on anatomical correctness, not model capacity or fitting artifacts, and that all models remain inference-independent. Cross-domain tests on echocardiography, cervical spine, and hand radiographs add weight to the claim that prior value scales with landmark spatial entropy.

The evidence lines up well. Five-fold CV (p=0.0015), patient-level permutation (p<0.0001), Grad-CAM activation differences (p<0.001), clinical kappa of 0.79-0.84 with no Class II/III reversals, and ICC >0.95 all converge on the same result. The 1% validation-to-test gap with priors versus 88% without is a clean demonstration.

The soft spot is the five-phase pipeline that generates the priors. No quantitative check (such as prior-to-GT landmark distance or overlap) is reported to confirm the priors are independent of the 25 ground-truth labels. If the pipeline encodes rules or features tied to the annotation process, the matrix gains could partly reflect leakage rather than pure regularization. That concern is real but not fatal given the random-prior control.

This paper is for people working on medical landmark detection who want a practical way to add domain knowledge without inference overhead. Readers focused on cephalometric analysis or similar tasks will find the ablation useful.

It deserves peer review. The central mechanism is tested with multiple converging checks, and the method is deployable. Send it forward, with a request to add a direct prior-GT independence metric in revision.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes a five-phase anatomy-guided pipeline to produce confidence-weighted spatial priors that encode clinical tracing workflows. These priors regularize HRNet-W32 training for 25 cephalometric landmarks, yielding 1.04 mm mean radial error on 1,502 radiographs from multiple devices. A training x inference prior matrix is used to isolate the contribution, showing that only image-specific anatomically correct priors (as opposed to random priors, no priors, or architecture changes) achieve the reported performance while maintaining a small 1% validation-to-test gap; the priors act solely as a training regularizer with no requirement at inference. Supporting analyses include five-fold CV (p=0.0015), patient-level permutation testing (p<0.0001), Grad-CAM activation differences (p<0.001), and clinical metrics (kappa 0.79-0.84, ICC>0.95). Cross-domain tests on echocardiography, cervical spine, and hand radiographs are presented to argue that prior utility scales with landmark spatial entropy.

Significance. If the priors can be shown to be independent of ground-truth labels, the work would offer a concrete mechanism for injecting anatomical domain knowledge into landmark detection training without inference overhead. The structured ablation matrix provides a useful template for disentangling prior effects from architecture or data leakage, and the clinical validation plus cross-domain results indicate potential applicability beyond cephalometry when landmark distributions exhibit high spatial entropy.

major comments (1)

[five-phase anatomy-guided pipeline and training x inference prior matrix] The central claim—that only image-specific anatomically correct priors produce the 1.04 mm result as a training-time regularizer—rests on the five-phase pipeline generating priors that accurately reflect true anatomy without systematic bias or label leakage from the 25 GT landmarks. No quantitative check (e.g., prior-mode to GT-landmark distance, spatial overlap, or correlation metrics) is reported in the pipeline description or the training x inference prior matrix analysis to confirm independence from the annotation process used for supervision. This leaves open the possibility that the performance gap versus random priors (1.72 mm) and the small val-to-test gap reflect partial leakage rather than pure anatomical regularization.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback, particularly the focus on confirming that the spatial priors are independent of ground-truth labels. We address the concern directly below.

read point-by-point responses

Referee: [five-phase anatomy-guided pipeline and training x inference prior matrix] The central claim—that only image-specific anatomically correct priors produce the 1.04 mm result as a training-time regularizer—rests on the five-phase pipeline generating priors that accurately reflect true anatomy without systematic bias or label leakage from the 25 GT landmarks. No quantitative check (e.g., prior-mode to GT-landmark distance, spatial overlap, or correlation metrics) is reported in the pipeline description or the training x inference prior matrix analysis to confirm independence from the annotation process used for supervision. This leaves open the possibility that the performance gap versus random priors (1.72 mm) and the small val-to-test gap reflect partial leakage rather than pure anatomical regularization.

Authors: We agree that a direct quantitative check for independence would strengthen the central claim. The five-phase pipeline derives priors from a structured clinical tracing workflow using general anatomical rules and image features, without reference to the specific 25 cephalometric ground-truth positions. The training × inference matrix already isolates the effect: only image-specific anatomically correct priors reach 1.04 mm with a 1% val-to-test gap, whereas random priors yield only 1.72 mm (unstable) and architecture-only or no-prior variants fail to close the gap. This differential performance is inconsistent with systematic leakage, which would be expected to advantage non-anatomical conditions similarly. Nevertheless, to address the referee's point explicitly, we will add the requested metrics (prior-mode to GT distance, spatial overlap, and correlation) in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

Derivation chain self-contained with no reduction to inputs by construction

full rationale

The paper establishes its central claim via an empirical training x inference prior matrix that directly compares anatomical priors against random priors, no priors, and architecture-only baselines, with the 1.04 mm result isolated to image-specific anatomical correctness. Supporting evidence includes five-fold cross-validation, patient-level permutation testing, Grad-CAM quantification, and clinical measurement validation (kappa, ICC), none of which reduce to the input labels by definition or self-citation. No equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text; the five-phase pipeline is described as an external anatomical workflow rather than a tautological encoding of the 25 GT landmarks. The result is therefore not forced by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the correctness of the five-phase pipeline for generating image-specific anatomical priors; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)

domain assumption The five-phase anatomy-guided pipeline produces image-specific anatomically correct priors that reflect true landmark distributions
Invoked throughout the pipeline description and the claim that only anatomically correct priors achieve 1.04 mm error.

pith-pipeline@v0.9.1-grok · 5796 in / 1214 out tokens · 38823 ms · 2026-07-01T00:38:01.861909+00:00 · methodology

Review history (3 revisions) →

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 3 canonical work pages · 1 internal anchor

[1]

H. W. Fields, B. E. Larson, D. M. Sarver, W. R. Proffit.Con- temporary Orthodontics, 7th ed. Elsevier, 2024

2024
[2]

Wang et al

C.-W. Wang et al. Evaluation and comparison of anatomical landmark detection methods for cephalometric x-ray images: A grand challenge.IEEE Trans. Med. Imaging, 34(9):1890–1900, 2015

1900
[3]

Lindner and T

C. Lindner and T. F. Cootes. Fully automatic cephalometric evaluation using random forest regression-voting. InProc. IEEE ISBI, 2015

2015
[4]

Zeng et al

M. Zeng et al. Cascaded convolutional networks for automatic cephalometric landmark detection.Medical Image Analysis, 68:101904, 2021

2021
[5]

Chen et al

R. Chen et al. Cephalometric landmark detection by attentive feature pyramid fusion and regression-voting. InProc. MICCAI, pp. 873–881, 2019

2019
[6]

Jaheen et al

A. Jaheen et al. CephRes-MHNet: A multi-head residual net- work for cephalometric landmark detection.arXiv:2511.10173, 2025

work page arXiv 2025
[7]

M. A. Khalid et al. CEPHA29: Automatic cephalometric land- mark detection challenge 2023.arXiv:2212.04808, 2022

work page arXiv 2023
[8]

H. J. Kwon et al. Automated cephalometric landmark detec- tion with confidence regions using Bayesian CNNs.BMC Oral Health, 20:270, 2020

2020
[9]

Son et al

I. Son et al. Ceph-Net: Automatic detection of cephalometric landmarks using an attention-based stacked regression network. BMC Oral Health, 2023

2023
[10]

Zhong et al

Z. Zhong et al. An attention-guided deep regression model for landmark detection in cephalograms. InProc. MICCAI, pp. 540–548, 2019

2019
[11]

Oh et al

K. Oh et al. Deep anatomical context feature learning for cephalometric landmark detection.IEEE J. Biomed. Health In- form., 2021

2021
[12]

M. A. Khalid et al. A two-stage regression framework for au- tomated cephalometric landmark detection.Expert Syst. Appl., 124840, 2024

2024
[13]

M. A. Khalid et al. A benchmark dataset for automatic cephalo- metric landmark detection.Scientific Data, 2025. 14

2025
[14]

R. R. Selvaraju et al. Grad-CAM: Visual explanations from deep networks via gradient-based localization. InProc. ICCV, pp. 618–626, 2017

2017
[15]

K. Sun, B. Xiao, D. Liu, and J. Wang. Deep high-resolution representation learning for visual recognition. InProc. CVPR, pp. 5693–5703, 2019

2019
[16]

Zhang, X

F. Zhang, X. Zhu, H. Dai, M. Ye, and C. Zhu. Distribution- aware coordinate representation for human pose estimation. In Proc. CVPR, pp. 7093–7102, 2020

2020
[17]

Payer, D

C. Payer, D. Štern, H. Bischof, and M. Urschler. Integrating spa- tial configuration into heatmap regression based CNNs for land- mark localization.Medical Image Analysis, 54:207–219, 2019

2019
[18]

Q. Ma, E. Kobayashi, and B. Fan. Automatic cephalometric landmark detection using modified Swin Transformer. InCL- Detection 2023 MICCAI Workshop, 2023

2023
[19]

Chen et al

L. Chen et al. CephalFormer: Multi-head attention in vision transformers for cephalometric landmark detection.Medical Image Analysis, 2023

2023
[20]

Wu et al

Y . Wu et al. Multi-scale feature fusion for cephalometric land- mark detection. InCL-Detection 2023 MICCAI Workshop, 2023

2023
[21]

Tian et al

Y . Tian et al. A comprehensive survey of cephalometric land- mark detection: Methods, datasets, and future directions.Artifi- cial Intelligence Review, 57:148, 2024

2024
[22]

Leclerc, E

S. Leclerc, E. Smistad, J. Pedrosa, A. Östvik, et al. Deep learning for segmentation using an open large-scale dataset in 2D echocardiography.IEEE Trans. Med. Imaging, 38(9):2198– 2210, 2019

2019
[23]

J. P. Howard et al. Automated left ventricular dimension assess- ment using artificial intelligence developed and validated by a UK-wide collaborative.Circulation: Cardiovascular Imaging, 14(5):e012135, 2021

2021
[24]

Ouyang et al

D. Ouyang et al. Video-based AI for beat-to-beat assessment of cardiac function.Nature, 580(7802):252–256, 2020

2020
[25]

K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into recti- fiers: Surpassing human-level performance on ImageNet classi- fication. InProc. ICCV, pp. 1026–1034, 2015

2015
[26]

Y . Ran, W. Qin, C. Qin, X. Li, Y . Liu, L. Xu, X. Mu, L. Yan, B. Wang, Y . Dai, J. Chen, and D. Han. A high-quality dataset featuring classified and annotated cervical spine X-ray atlas. Scientific Data, 11(1):631, 2024

2024
[27]

Ronneberger, P

O. Ronneberger, P. Fischer, and T. Brox. U-Net: Convolutional networks for biomedical image segmentation. InProc. MICCAI, pp. 234–241, 2015

2015
[28]

Sandler, A

M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.- C. Chen. MobileNetV2: Inverted residuals and linear bottle- necks. InProc. CVPR, pp. 4510–4520, 2018

2018
[29]

K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InProc. CVPR, pp. 770–778, 2016

2016
[30]

Distilling the Knowledge in a Neural Network

G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network.arXiv:1503.02531, 2015

work page internal anchor Pith review Pith/arXiv arXiv 2015
[31]

Loshchilov and F

I. Loshchilov and F. Hutter. Decoupled weight decay regular- ization. InProc. ICLR, 2019

2019
[32]

Loshchilov and F

I. Loshchilov and F. Hutter. SGDR: Stochastic gradient descent with warm restarts. InProc. ICLR, 2017

2017
[33]

D. H. Douglas and T. K. Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature.Cartographica, 10(2):112–122, 1973

1973
[34]

bright band

A. Gertych, A. Zhang, J. Sayre, S. Pospiech-Kurkowska, and H. K. Huang. Bone age assessment of children using a digi- tal hand atlas.Computerized Medical Imaging and Graphics, 31(4–5):322–331, 2007. 15 Supplementary Materials: Tracing Like a Clinician: Anatomy-Guided Spatial Priors for Cephalometric Landmark Detection Sidhartha Mohapatra1, Dr. Pallavi Moh...

2007

[1] [1]

H. W. Fields, B. E. Larson, D. M. Sarver, W. R. Proffit.Con- temporary Orthodontics, 7th ed. Elsevier, 2024

2024

[2] [2]

Wang et al

C.-W. Wang et al. Evaluation and comparison of anatomical landmark detection methods for cephalometric x-ray images: A grand challenge.IEEE Trans. Med. Imaging, 34(9):1890–1900, 2015

1900

[3] [3]

Lindner and T

C. Lindner and T. F. Cootes. Fully automatic cephalometric evaluation using random forest regression-voting. InProc. IEEE ISBI, 2015

2015

[4] [4]

Zeng et al

M. Zeng et al. Cascaded convolutional networks for automatic cephalometric landmark detection.Medical Image Analysis, 68:101904, 2021

2021

[5] [5]

Chen et al

R. Chen et al. Cephalometric landmark detection by attentive feature pyramid fusion and regression-voting. InProc. MICCAI, pp. 873–881, 2019

2019

[6] [6]

Jaheen et al

A. Jaheen et al. CephRes-MHNet: A multi-head residual net- work for cephalometric landmark detection.arXiv:2511.10173, 2025

work page arXiv 2025

[7] [7]

M. A. Khalid et al. CEPHA29: Automatic cephalometric land- mark detection challenge 2023.arXiv:2212.04808, 2022

work page arXiv 2023

[8] [8]

H. J. Kwon et al. Automated cephalometric landmark detec- tion with confidence regions using Bayesian CNNs.BMC Oral Health, 20:270, 2020

2020

[9] [9]

Son et al

I. Son et al. Ceph-Net: Automatic detection of cephalometric landmarks using an attention-based stacked regression network. BMC Oral Health, 2023

2023

[10] [10]

Zhong et al

Z. Zhong et al. An attention-guided deep regression model for landmark detection in cephalograms. InProc. MICCAI, pp. 540–548, 2019

2019

[11] [11]

Oh et al

K. Oh et al. Deep anatomical context feature learning for cephalometric landmark detection.IEEE J. Biomed. Health In- form., 2021

2021

[12] [12]

M. A. Khalid et al. A two-stage regression framework for au- tomated cephalometric landmark detection.Expert Syst. Appl., 124840, 2024

2024

[13] [13]

M. A. Khalid et al. A benchmark dataset for automatic cephalo- metric landmark detection.Scientific Data, 2025. 14

2025

[14] [14]

R. R. Selvaraju et al. Grad-CAM: Visual explanations from deep networks via gradient-based localization. InProc. ICCV, pp. 618–626, 2017

2017

[15] [15]

K. Sun, B. Xiao, D. Liu, and J. Wang. Deep high-resolution representation learning for visual recognition. InProc. CVPR, pp. 5693–5703, 2019

2019

[16] [16]

Zhang, X

F. Zhang, X. Zhu, H. Dai, M. Ye, and C. Zhu. Distribution- aware coordinate representation for human pose estimation. In Proc. CVPR, pp. 7093–7102, 2020

2020

[17] [17]

Payer, D

C. Payer, D. Štern, H. Bischof, and M. Urschler. Integrating spa- tial configuration into heatmap regression based CNNs for land- mark localization.Medical Image Analysis, 54:207–219, 2019

2019

[18] [18]

Q. Ma, E. Kobayashi, and B. Fan. Automatic cephalometric landmark detection using modified Swin Transformer. InCL- Detection 2023 MICCAI Workshop, 2023

2023

[19] [19]

Chen et al

L. Chen et al. CephalFormer: Multi-head attention in vision transformers for cephalometric landmark detection.Medical Image Analysis, 2023

2023

[20] [20]

Wu et al

Y . Wu et al. Multi-scale feature fusion for cephalometric land- mark detection. InCL-Detection 2023 MICCAI Workshop, 2023

2023

[21] [21]

Tian et al

Y . Tian et al. A comprehensive survey of cephalometric land- mark detection: Methods, datasets, and future directions.Artifi- cial Intelligence Review, 57:148, 2024

2024

[22] [22]

Leclerc, E

S. Leclerc, E. Smistad, J. Pedrosa, A. Östvik, et al. Deep learning for segmentation using an open large-scale dataset in 2D echocardiography.IEEE Trans. Med. Imaging, 38(9):2198– 2210, 2019

2019

[23] [23]

J. P. Howard et al. Automated left ventricular dimension assess- ment using artificial intelligence developed and validated by a UK-wide collaborative.Circulation: Cardiovascular Imaging, 14(5):e012135, 2021

2021

[24] [24]

Ouyang et al

D. Ouyang et al. Video-based AI for beat-to-beat assessment of cardiac function.Nature, 580(7802):252–256, 2020

2020

[25] [25]

K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into recti- fiers: Surpassing human-level performance on ImageNet classi- fication. InProc. ICCV, pp. 1026–1034, 2015

2015

[26] [26]

Y . Ran, W. Qin, C. Qin, X. Li, Y . Liu, L. Xu, X. Mu, L. Yan, B. Wang, Y . Dai, J. Chen, and D. Han. A high-quality dataset featuring classified and annotated cervical spine X-ray atlas. Scientific Data, 11(1):631, 2024

2024

[27] [27]

Ronneberger, P

O. Ronneberger, P. Fischer, and T. Brox. U-Net: Convolutional networks for biomedical image segmentation. InProc. MICCAI, pp. 234–241, 2015

2015

[28] [28]

Sandler, A

M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.- C. Chen. MobileNetV2: Inverted residuals and linear bottle- necks. InProc. CVPR, pp. 4510–4520, 2018

2018

[29] [29]

K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InProc. CVPR, pp. 770–778, 2016

2016

[30] [30]

Distilling the Knowledge in a Neural Network

G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network.arXiv:1503.02531, 2015

work page internal anchor Pith review Pith/arXiv arXiv 2015

[31] [31]

Loshchilov and F

I. Loshchilov and F. Hutter. Decoupled weight decay regular- ization. InProc. ICLR, 2019

2019

[32] [32]

Loshchilov and F

I. Loshchilov and F. Hutter. SGDR: Stochastic gradient descent with warm restarts. InProc. ICLR, 2017

2017

[33] [33]

D. H. Douglas and T. K. Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature.Cartographica, 10(2):112–122, 1973

1973

[34] [34]

bright band

A. Gertych, A. Zhang, J. Sayre, S. Pospiech-Kurkowska, and H. K. Huang. Bone age assessment of children using a digi- tal hand atlas.Computerized Medical Imaging and Graphics, 31(4–5):322–331, 2007. 15 Supplementary Materials: Tracing Like a Clinician: Anatomy-Guided Spatial Priors for Cephalometric Landmark Detection Sidhartha Mohapatra1, Dr. Pallavi Moh...

2007