A Hierarchical Slice Attention Network for Appendicitis Classification in 3D CT Scans

Chia-Wen Huang; Chien-Chang Lee; Haw Hwai; Pei-Yuan Wu

REVIEW 4 major objections 5 minor 22 references

A Hierarchical Slice Attention Network for Appendicitis Classification in 3D CT Scans

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A 2D-guided slice attention network improves appendicitis AUC by 3% over prior work, reaching 0.8887 on internal validation.

desk verdict A plausible but not fully pinned-down AUC improvement for appendicitis classification; the main new ingredient (external 2D slice supervision) shows only a small effect, and the preprocessing window is underspecified. read the letter →

arxiv 2506.23209 v1 pith:OUM5VXFK submitted 2025-06-29 cs.CV

Chia-Wen Huang , Haw Hwai , Chien-Chang Lee , Pei-Yuan Wu This is my paper · ORCID

classification cs.CV

keywords appendicitisclassification3DCTsliceattentionhierarchicalSwinTransformertransferlearningcoherencelossmedicalimageanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a hierarchical slice attention network can classify appendicitis from 3D CT scans using mostly 2D components, with external 2D slice-level supervision steering which slices matter. The model first decides whether appendicitis is present, then whether it is simple or complicated, and it weights each of 64 axial slices by relevance instead of requiring a radiologist to draw a region of interest. On internal validation it reports AUC 0.8887 for appendicitis and 0.8833 for complicated appendicitis, beating the previous appendicitis-specific baseline, and on external validation the complicated-appendicitis AUC stays at 0.8803. The clinical point is that distinguishing simple from complicated appendicitis changes treatment, so an automated model that does both in one pass could help triage while radiologist workloads grow.

What carries the argument

The central mechanism is the Slice Attention Alignment Module, which enforces a coherence loss between the attention weights computed from 3D slice features and inverse-distance weights computed from learned 2D class centers. A 2D Swin Transformer processes each of 64 axial slices; an auxiliary branch trained on external slice-labeled data learns positive and negative class centers through a Center Loss; and the distance from each slice feature to the positive center is converted into an alignment weight. A hierarchical classifier, following the Local Classifier per Level scheme, first aggregates slice features for appendicitis probability and then concatenates the aggregate with each slice feature to decide simple versus complicated appendicitis. This alignment is what lets slice-level 2D knowledge guide the 3D classification without manual region-of-interest selection.

What would settle it

Take a held-out set of CT scans and split them by where the appendix lies within the chosen 64-slice window: if cases near the window boundary show a large drop in AUC relative to central cases, the fixed-window assumption is doing the work. A second check would be to compare the model's highest-attention slices against radiologist-identified appendix locations; low overlap would show the attention is not actually finding the appendix.

Watch

Extended reading notes

Core claim

The central claim is that adding external 2D datasets with slice-level annotations, plus a hierarchical classification stage, improves appendicitis classification over prior work: AUC rises by 3% for appendicitis and 5.9% for complicated appendicitis. Concretely, the model reaches AUC 0.8887 for appendicitis and 0.8833 for complicated appendicitis in internal validation, compared with 0.8585 and 0.8244 for the appendix-specific baseline and 0.7901 and 0.7846 for a generic 3D DenseNet. In external validation, where the model trains on one hospital dataset and tests on another, the complicated-appendicitis AUC is 0.8803. The authors attribute the gains to pretraining a 2D Swin Transformer on radiology images, learning class centers from slice-labeled external data, aligning slice attention to those centers, and splitting the task into appendicitis detection and type classification.

Load-bearing premise

The load-bearing premise is that the appendix always falls inside the 64 axial slices selected by centering each CT scan at a fixed index and cropping the right lower abdomen, so the model never has to localize the appendix on its own.

Editorial extensions

If this is right

If the reported gains hold, abdominal CT classifiers can be assembled from 2D pretrained backbones plus slice attention instead of requiring large annotated 3D datasets.
The hierarchy means one model outputs both a presence probability and a severity class, which maps directly onto triage and surgical-urgency decisions.
Slice-level external supervision could let hospitals with only whole-scan labels benefit from datasets that are annotated slice by slice.
Keeping sensitivity fixed while improving specificity means the model reduces false positives without increasing missed diagnoses, a property that matters in emergency settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial extension: the coherence-loss mechanism is not appendicitis-specific; the same alignment of 2D slice-level supervision with 3D slice attention could transfer to other small-lesion volumetric tasks such as kidney stones or colitis.
Editorial extension: the strong external-validation result for complicated appendicitis suggests the auxiliary 2D supervision may act as a partial domain-transfer bridge between hospitals, but the paper does not isolate that effect as a measured claim.
Editorial extension: the attention weights could be read as weak localization of the appendix, yet the paper reports no spatial accuracy evaluation; comparing attended slices against radiologist-identified appendix locations would be a natural next test.
Editorial extension: the fixed 64-slice centering rule is a testable limitation; replacing it with an anatomical localizer would be a direct follow-up, and the current numbers should not be assumed to hold for scans with different coverage.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Quick take on arXiv:2506.23209. It's a competent, clearly written medical imaging paper: 3D CT appendicitis classification using a 2D Swin Transformer per slice, slice attention, a hierarchical simple/complicated split, and an auxiliary 2D module with center loss and coherence loss guided by an external slice-labeled dataset (AOCR2024). The genuinely new piece is applying that Sun et al. breast-ultrasound keyframe-center idea to appendicitis and combining it with hierarchical classification. The authors also did a two-hospital external validation and took care to exclude overlapping cases between FEMH and AOCR2024, which is more than many papers do.

What it does well: the ablation is honest. It shows the baseline 2D Swin gets AUC 0.8674 for appendicitis; adding the hierarchy gets 0.8847; adding the 2D module gets 0.8887. So the advertised "external 2D guidance" contributes only ~0.004 AUC for appendicitis and ~0.003 for complicated. The hierarchy is doing most of the work. The paper does not oversell this, but the abstract's "improves AUC by 3%" is relative to AppendixNet, not to the ablation.

Where I'd push back:

- No confidence intervals, no significance tests anywhere. The headline improvements over AppendixNet on external validation (0.7413 vs 0.7044 for appendicitis; 0.8803 vs 0.7601 for complicated) look meaningful, but with no error bars I can't tell if the appendicitis gap is noise. This is the single biggest fix needed. - The preprocessing is under-specified. Section 3.1.2 says scans were "centered around a fixed index" to pick 64 slices, and a right-lower-abdomen crop gets resized to 224x224. They never say what that index is, how it was chosen, or whether the appendix was verified to be inside the window. If the appendix falls outside for some patients, the model can't see the target. All compared methods share the same preprocessing, so the relative comparison is fair, but the absolute AUC claims for a "reliable diagnostic tool" depend on this window being sensible. - The 3D DenseNet baseline is embarrassing: external AUC 0.4496 for appendicitis, below chance. That makes the comparison to their model look better than it should. They should either train it better or drop it and compare against AppendixNet and maybe a 3D ResNet with honest tuning. - Thresholds were tuned to hit sensitivity 0.9/0.8, so specificity comparisons are conditional on that post-hoc choice. AUC is threshold-independent, so the main claim survives, but the practical sensitivity/specificity numbers should be read with that caveat.

Overall: the hierarchical framework looks sound and the external validation is a real strength. The 2D-guidance contribution is smaller than the framing suggests, and the missing error bars and underspecified slice window keep this from being definitive. It deserves peer review; an editor should send it out, and referees should ask for CIs, a clearer preprocessing description, and a stronger 3D baseline.

Would I cite it? Probably not in the next 12 months unless the CIs show the external gap holds. But I'd bring it to reading group—there's a useful discussion about what counts as a "contribution" when the headline module adds 0.004 AUC.

That's my read.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes a hierarchical model for classifying appendicitis and its simple/complicated status from 3D abdominal CT scans. A 2D Swin Transformer pretrained on RadImageNet processes 64 axial slices per scan; an auxiliary module trained on slice-annotated 2D data (AOCR2024) learns class centers, and a coherence loss aligns the 3D slice-attention weights with distances to those centers. The model is evaluated on two hospital datasets (NTUH and FEMH) with internal (mixed) and external (FEMH-to-NTUH) validation, and is compared with AppendixNet and a 3D DenseNet baseline, followed by ablations. The authors report AUC improvements of 3% for appendicitis and 5.9% for complicated appendicitis over previous work.

Significance. Appendicitis diagnosis from CT is clinically important, and the proposed architecture makes a plausible contribution by transferring 2D pretrained knowledge into 3D classification while distinguishing simple from complicated appendicitis. The external validation and the ablation study are useful elements, and the use of slice-level external supervision is a reasonable way to compensate for scarce 3D labels. If the reported improvements are statistically reliable, the method would be a practical step toward automated appendicitis triage. However, the paper does not provide confidence intervals or significance tests, one external-validation baseline is below chance, and the input preprocessing is not sufficiently described or validated, so the magnitude and robustness of the claimed advantages are not yet established.

major comments (4)

[3.1.2] The input construction is not adequately specified or validated. The paper states that each patient's CT scan 'was centered around a fixed index, selecting 64 slices in total' and that '2D slices from the right lower abdomen were resized to 224×224 pixels,' but it never identifies the fixed index, how it was chosen, or the anatomical coverage of the resulting window. Since the appendix must lie inside this 64-slice, 224×224 window for the model to see it, the authors should report the fixed index, its anatomical correlate, and the distribution of appendix positions within the window across both hospitals. If the window was selected by inspecting the data, the internal evaluation is biased; if it is not anatomically anchored, the external 'reliable diagnostic solution' claim is unsupported even though all compared models share the same preprocessing.
[3.2, Tables 1 and 2] The central claim of improved AUC is not accompanied by any statistical uncertainty quantification. Differences such as 0.8887 versus 0.8585 (internal appendicitis AUC) and 0.7413 versus 0.7044 (external appendicitis AUC) could be within sampling noise given the test sizes; no confidence intervals (e.g., DeLong or bootstrap), paired significance tests, or multiple-seed variances are reported. Because the operating thresholds were adjusted to hit fixed sensitivities, the specificity gains shown in Tables 1 and 2 are threshold-dependent and should be summarized with ROC curves or partial AUCs. These additions are needed to support the paper's quantitative superiority claims.
[Table 2] The external-validation result for 3D DenseNet on appendicitis (AUC 0.4496, specificity 0.1050) is below chance, which indicates that the baseline is not behaving as expected under the external split. Even though all models share preprocessing, this anomaly suggests a training or distribution-shift problem in the baseline rather than a meaningful comparison. The authors should diagnose whether this is due to hyperparameters, class imbalance, or the external evaluation protocol, and should report baseline training details; as presented, Table 2 does not support a fair comparison for that baseline.
[Table 3] The ablation increments are small (e.g., from 0.8847 to 0.8887 for appendicitis AUC and from 0.8804 to 0.8833 for complicated appendicitis AUC) and no uncertainty estimates are given, so the study does not demonstrate that the 2D module and coherence loss contribute beyond the hierarchical baseline. Reporting seed-averaged results or interval estimates, and ideally attention maps or a quantitative analysis of selected slices, would substantiate the claim that external 2D supervision improves key-slice selection.

minor comments (5)

[3.1.2] The threshold tuning described as 'adjusted to achieve a sensitivity of 0.9 for appendicitis and 0.8 for complicated appendicitis' is reported only qualitatively; the paper should state the actual threshold values and present results at multiple operating points.
[2.2, Eq. (4)] Equation (4) reuses N for what appears to be a batch of 2D slices while N=64 was defined for the number of 3D slices; clarifying the notation would avoid ambiguity.
[3.1.2] The DICOM window center and width settings are said to be fixed but are not reported; without these values, the preprocessing is not fully reproducible.
[References] Reference [11] is cited with a '0' volume/issue placeholder; please provide the final citation details.
[Figure 1] Figure 1 is not referenced in the body text except as a general overview; consider referring to it when introducing the Main Module, Auxiliary Module, and Slice Attention Alignment Module.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported AUC improvements are empirical results from held-out test sets, with no derivation that reduces to fitted inputs or self-citations.

full rationale

This paper makes no formal derivation claim; it reports an empirical comparison of deep learning models on internal and external validation sets. The central claims are the AUC values in Tables 1 and 2, which are computed on held-out test data (15% split, with external validation testing on NTUH after training on FEMH). No model parameter is fitted to the test labels and then renamed as a prediction. The sensitivity values are matched to fixed targets (0.9 for appendicitis, 0.8 for complicated appendicitis) by threshold adjustment, but this does not affect the threshold-independent AUC metric and is standard practice. The auxiliary 2D supervision uses AOCR2024, which is sourced from FEMH, but the authors explicitly state that overlapping cases were excluded from AOCR2024 to preserve the integrity of the FEMH test set; this is a data-hygiene measure, not a circular construction. The cited prior works, including Sun et al. [7] for coherence loss and Rajpurkar et al. [4] for AppendixNet, are external and not authored by the present authors, so no load-bearing self-citation exists. The unspecified fixed slice index and right-lower-abdomen crop are potential validity concerns about whether the appendix is always inside the input window, but that is a correctness and generalization risk, not circularity. The paper's conclusions are supported by independent test-set evaluations, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper relies on standard deep learning practices and does not introduce new theoretical entities. The main assumptions are about anatomy (appendix location) and transferability of 2D features to 3D slices.

free parameters (4)

Loss weights (alpha, beta, gamma, delta, lambda) = not reported
Weights in the total loss (Eq. 6) are not given; they are hand-chosen or tuned and affect the balance among tasks.
Classification thresholds = sensitivity 0.9 (appendicitis), 0.8 (complicated)
Thresholds are adjusted to achieve target sensitivities, likely on the validation or test set, directly affecting reported specificity and accuracy.
Slice count N = 64
Each scan is reduced to 64 slices; the choice is ad hoc and not justified by appendix anatomy.
Fixed centering index = not specified
The reference slice used to center the 64-slice window is not described; it determines whether the appendix is included.

assumptions (3)

domain assumption The appendix is located within the selected 64 axial slices and the right-lower-abdomen crop used as network input.
Section 3.1.2 states all scans are standardized by centering around a fixed index and resizing right lower abdomen slices; if the appendix is outside this window, classification is impossible.
domain assumption External 2D slices from AOCR2024 with slice-level labels are representative of the 3D CT slices and share the same feature distribution when processed by the shared Swin encoder.
The coherence loss (Eq. 5) aligns 3D attention weights with distances to 2D-learned centers; this assumes the Swin features are comparable across the two datasets despite different labeling granularity and acquisition.
domain assumption The hierarchical classification pipeline (appendicitis first, then type) does not accumulate errors, i.e., the type classifier is only evaluated on true appendicitis cases or the loss handles negative cases correctly.
The method trains L_app and L_type jointly but does not specify how type prediction is produced or evaluated when appendicitis is absent; this affects the reported complicated appendicitis metrics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Hierarchical Slice Attention Network for Appendicitis Classification in 3D CT Scans." pith.science (2026). https://pith.science/paper/OUM5VXFK

@misc{pith2026250623209,
  author       = {Pith},
  title        = {Pith review of: A Hierarchical Slice Attention Network for Appendicitis Classification in 3D CT Scans},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OUM5VXFK}},
  note         = {Machine review of arXiv:2506.23209}
}

read the original abstract

Timely and accurate diagnosis of appendicitis is critical in clinical settings to prevent serious complications. While CT imaging remains the standard diagnostic tool, the growing number of cases can overwhelm radiologists, potentially causing delays. In this paper, we propose a deep learning model that leverages 3D CT scans for appendicitis classification, incorporating Slice Attention mechanisms guided by external 2D datasets to enhance small lesion detection. Additionally, we introduce a hierarchical classification framework using pre-trained 2D models to differentiate between simple and complicated appendicitis. Our approach improves AUC by 3% for appendicitis and 5.9% for complicated appendicitis, offering a more efficient and reliable diagnostic solution compared to previous work.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages

[1]

Computed tomography (CT) imaging plays a key role in diagnosing appendicitis and re- lated conditions

INTRODUCTION Accurate and timely diagnosis of appendicitis is crucial to preventing serious conditions. Computed tomography (CT) imaging plays a key role in diagnosing appendicitis and re- lated conditions. However, the increasing number of cases can overwhelm radiologists, impacting their diagnostic ca- pacity. To address this, automated diagnostic tools...

work page
[2]

We propose a deep learning model for appendicitis classification using 3D CT scans, incorporating exter- nal 2D datasets with slice-level annotations to automate key slice identification and improves accuracy

work page
[3]

We introduce a hierarchical classification framework that differentiates between simple and complicated ap- pendicitis, providing more detailed diagnostic insights to support clinical decision-making

work page
[4]

A Hierarchical Slice Attention Network for Appendicitis Classification in 3D CT Scans

We demonstrate that integrating pre-trained 2D models into 3D CT tasks improves classification accuracy for both appendicitis and complicated appendicitis. arXiv:2506.23209v1 [cs.CV] 29 Jun 2025 Fig. 1. Overview of the proposed model. The framework includes: (a) the Auxiliary Module, which processes external 2D data with a 2D Swin Transformer; (b) the Mai...

work page Pith review arXiv 2025
[5]

METHODS Our framework classifies 3D CT images for appendicitis and complicated appendicitis by integrating both 2D and 3D in- formation, as shown in Fig. 1. 2.1. Main Module: Hierarchical Classification Frame- work We adopt a hierarchical classification (HC) approach, follow- ing the Local Classifier per Level method outlined in [8]. The framework first c...

work page
[6]

Total Loss Function The total loss function combines the losses from both the 2D and 3D tasks: Ltotal = α · Lapp + β · Ltype + γ · Lcenter + δ · Lcoherence + λ · L2D

(5) 2.4. Total Loss Function The total loss function combines the losses from both the 2D and 3D tasks: Ltotal = α · Lapp + β · Ltype + γ · Lcenter + δ · Lcoherence + λ · L2D. (6)

work page
[7]

Dataset and Model Implementation 3.1.1

EXPERIMENTS 3.1. Dataset and Model Implementation 3.1.1. Dataset Overview This study utilized appendicitis-related datasets from two hospitals: NTUH (National Taiwan University Hospital) and FEMH (Far Eastern Memorial Hospital). Both datasets in- clude cases of patients who underwent contrast-enhanced CT scans for suspected appendicitis. Pregnant patients...

work page
[8]

Leveraging Slice Attention and external 2D datasets enhances small lesion classification, and integrating pre-trained 2D models improves accuracy and robustness

CONCLUSION In conclusion, we propose a 2D hierarchical deep learning model for classifying appendicitis and complicated appen- dicitis from 3D CT scans. Leveraging Slice Attention and external 2D datasets enhances small lesion classification, and integrating pre-trained 2D models improves accuracy and robustness. Our approach effectively distinguishes be-...

work page

Show all 22 references

[9]

Due to the retrospective nature of the study, informed consent was waived

COMPLIANCE WITH ETHICAL STANDARDS This retrospective study was conducted following the Decla- ration of Helsinki and was approved by the Research Ethics Committee of National Taiwan University Hospital. Due to the retrospective nature of the study, informed consent was waived....
[10]

Deeplesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning,

Ke Yan, Xiaosong Wang, Le Lu, and Ronald M Sum- mers, “Deeplesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning,” Journal of medical imaging, vol. 5, no. 3, pp. 036501–036501, 2018

2018
[11]

The brain tumor seg- mentation (brats) challenge 2023: Focus on pediatrics (cbtn-connect-dipgr-asnr-miccai brats-peds),

Anahita Fathi Kazerooni, Nastaran Khalili, Xinyang Liu, Debanjan Haldar, Zhifan Jiang, Syed Muhammed Anwar, Jake Albrecht, Maruf Adewole, Udunna Ana- zodo, Hannah Anderson, et al., “The brain tumor seg- mentation (brats) challenge 2023: Focus on pediatrics (cbtn-connect-dipgr-...

2023 arXiv
[12]

Crowdsourcing pneumotho- rax annotations using machine learning annotations on the nih chest x-ray dataset,

Ross W Filice, Anouk Stein, Carol C Wu, Veron- ica A Arteaga, Stephen Borstelmann, Ramya Gaddikeri, Maya Galperin-Aizenberg, Ritu R Gill, Myrna C Godoy, Stephen B Hobbs, et al., “Crowdsourcing pneumotho- rax annotations using machine learning annotations on the nih chest x-ray...

2020
[13]

Appendixnet: deep learning for diag- nosis of appendicitis from a small dataset of ct exams using video pretraining,

Pranav Rajpurkar, Allison Park, Jeremy Irvin, Chris Chute, Michael Bereket, Domenico Mastrodicasa, Cur- tis P Langlotz, Matthew P Lungren, Andrew Y Ng, and Bhavik N Patel, “Appendixnet: deep learning for diag- nosis of appendicitis from a small dataset of ct exams using video ...

2020
[14]

Convolutional-neural-network-based diagnosis of appendicitis via ct scans in patients with acute ab- dominal pain presenting in the emergency department,

Jin Joo Park, Kyung Ah Kim, Yoonho Nam, Moon Hyung Choi, Sun Young Choi, and Jeongbae Rhie, “Convolutional-neural-network-based diagnosis of appendicitis via ct scans in patients with acute ab- dominal pain presenting in the emergency department,” Scientific reports, vol. 10, ...

2020
[15]

So Hyun Park, Young Jae Kim, Kwang Gi Kim, Jun- Won Chung, Hyun Cheol Kim, In Young Choi, Myung- Won You, Gi Pyo Lee, and Jung Han Hwang, “Com- parison between single and serial computed tomogra- phy images in classification of acute appendicitis, acute right-sided diverticuli...

2023
[16]

Boosting breast ultrasound video classification by the guidance of keyframe fea- ture centers,

Anlan Sun, Zhao Zhang, Meng Lei, Yuting Dai, Dong Wang, and Liwei Wang, “Boosting breast ultrasound video classification by the guidance of keyframe fea- ture centers,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2023, p...

2023
[17]

A survey of hi- erarchical classification across different application do- mains,

Carlos N Silla and Alex A Freitas, “A survey of hi- erarchical classification across different application do- mains,” Data mining and knowledge discovery, vol. 22, pp. 31–72, 2011

2011
[18]

Holis- tic and comprehensive annotation of clinically signifi- cant findings on diverse ct images: learning from radi- ology reports and label ontology,

Ke Yan, Yifan Peng, Veit Sandfort, Mohammadhadi Bagheri, Zhiyong Lu, and Ronald M Summers, “Holis- tic and comprehensive annotation of clinically signifi- cant findings on diverse ct images: learning from radi- ology reports and label ontology,” in Proceedings of the IEEE/CVF ...

2019
[19]

Learning hierarchical attention for weakly-supervised chest x-ray abnormality localization and diagnosis,

Xi Ouyang, Srikrishna Karanam, Ziyan Wu, Terrence Chen, Jiayu Huo, Xiang Sean Zhou, Qian Wang, and Jie-Zhi Cheng, “Learning hierarchical attention for weakly-supervised chest x-ray abnormality localization and diagnosis,” IEEE transactions on medical imaging, vol. 40, no. 10, ...

2020
[20]

Radimagenet: An open radio- logic deep learning research dataset for effective transfer learning,

Xueyan Mei, Zelong Liu, Philip M. Robson, Brett Marinelli, Mingqian Huang, Amish Doshi, Adam Ja- cobi, Chendi Cao, Katherine E. Link, Thomas Yang, Ying Wang, Hayit Greenspan, Timothy Deyer, Zahi A. Fayad, and Yang Yang, “Radimagenet: An open radio- logic deep learning research...
[21]

Swin transformer: Hierarchical vision transformer us- ing shifted windows,

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo, “Swin transformer: Hierarchical vision transformer us- ing shifted windows,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 10012–10022

2021
[22]

Aocr2024 ai challenge,

Wen-Jeng Lee Goman, Taiwan Radiological Soci- ety (TRS), “Aocr2024 ai challenge,” 2023

2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Computed tomography (CT) imaging plays a key role in diagnosing appendicitis and re- lated conditions

INTRODUCTION Accurate and timely diagnosis of appendicitis is crucial to preventing serious conditions. Computed tomography (CT) imaging plays a key role in diagnosing appendicitis and re- lated conditions. However, the increasing number of cases can overwhelm radiologists, impacting their diagnostic ca- pacity. To address this, automated diagnostic tools...

work page

[2] [2]

We propose a deep learning model for appendicitis classification using 3D CT scans, incorporating exter- nal 2D datasets with slice-level annotations to automate key slice identification and improves accuracy

work page

[3] [3]

We introduce a hierarchical classification framework that differentiates between simple and complicated ap- pendicitis, providing more detailed diagnostic insights to support clinical decision-making

work page

[4] [4]

A Hierarchical Slice Attention Network for Appendicitis Classification in 3D CT Scans

We demonstrate that integrating pre-trained 2D models into 3D CT tasks improves classification accuracy for both appendicitis and complicated appendicitis. arXiv:2506.23209v1 [cs.CV] 29 Jun 2025 Fig. 1. Overview of the proposed model. The framework includes: (a) the Auxiliary Module, which processes external 2D data with a 2D Swin Transformer; (b) the Mai...

work page Pith review arXiv 2025

[5] [5]

METHODS Our framework classifies 3D CT images for appendicitis and complicated appendicitis by integrating both 2D and 3D in- formation, as shown in Fig. 1. 2.1. Main Module: Hierarchical Classification Frame- work We adopt a hierarchical classification (HC) approach, follow- ing the Local Classifier per Level method outlined in [8]. The framework first c...

work page

[6] [6]

Total Loss Function The total loss function combines the losses from both the 2D and 3D tasks: Ltotal = α · Lapp + β · Ltype + γ · Lcenter + δ · Lcoherence + λ · L2D

(5) 2.4. Total Loss Function The total loss function combines the losses from both the 2D and 3D tasks: Ltotal = α · Lapp + β · Ltype + γ · Lcenter + δ · Lcoherence + λ · L2D. (6)

work page

[7] [7]

Dataset and Model Implementation 3.1.1

EXPERIMENTS 3.1. Dataset and Model Implementation 3.1.1. Dataset Overview This study utilized appendicitis-related datasets from two hospitals: NTUH (National Taiwan University Hospital) and FEMH (Far Eastern Memorial Hospital). Both datasets in- clude cases of patients who underwent contrast-enhanced CT scans for suspected appendicitis. Pregnant patients...

work page

[8] [8]

Leveraging Slice Attention and external 2D datasets enhances small lesion classification, and integrating pre-trained 2D models improves accuracy and robustness

CONCLUSION In conclusion, we propose a 2D hierarchical deep learning model for classifying appendicitis and complicated appen- dicitis from 3D CT scans. Leveraging Slice Attention and external 2D datasets enhances small lesion classification, and integrating pre-trained 2D models improves accuracy and robustness. Our approach effectively distinguishes be-...

work page