Rigor Reviewed

Rigor-Reviewed Manuscript

Sign language recognition using the fusion of image and hand landmarks through multi-headed convolutional neural network

Name: Αlpha¹ Rigor Record — Sign language recognition using the fusion of image and hand landmarks through multi-headed convolutional neural network
Item: Sign language recognition using the fusion of image and hand landmarks through multi-headed convolutional neural network
Rating: 70
Author: Αlpha¹

Pathan RK, Biswas M, Yasmin S, Khandaker MU, Salman M, Youssef AAF.

Paper source

Read full text (PMC)Journal

Embed this Rigor badge

Αlpha¹ Rigor Review — 3.25★ (Adequate), 0 critical, 5 high issues for “Sign language recognition using the fusion of image and hand landmarks through multi-headed convolutional neural network”

Links back to this Rigor Record.

Αlpha¹ Editorial

This rigor review was examined and confirmed by Αlpha¹ Editorial · August 5, 2026

Αlpha¹ Rigor Review

Rigor Reviewed

3.25 / 5

Adequate — real issues to address

Computational Study

Critical

High-priority

How this rating was calculated▸

Started at 5★ — deductions

IntegrityIntegrity concern−0.5★
ReportingKey resources not met−0.5★
LinksDead data/code link ×2−0.5★
ReportingStudy design partially met−0.25★

Model-reported confidence: moderateReviewer agreement: lowEngine: 7.15.0Source: full text

Not verified in this report

Statistics were not checked: no recomputable values were found in this text — no test statistic reported with its degrees of freedom, no effect estimate printed with both a 95% CI and a p-value, and no percentage printed with both its count and its denominator.

Problems surfaced4 types

Key resources not identified
Dead data/code links
Other integrity concern
Study-design details incomplete (controls, blinding, power)

This Αlpha¹ Rigor Review uses AI Rigor Reviewers trained on a curated corpus of high-fidelity and retracted papers, with expert supervision and curation. It can still make mistakes; verify each finding against the source before relying on it.

The paper presents a sensible computational approach for ASL recognition with a multi-headed CNN, but has several reporting gaps: software tools lack version numbers, no code is shared, the data URL is dead, and there is no conflict of interest statement. The text contains grammatical errors and a tortured phrase ('convolutional neural organization') that raises concerns about potential language manipulation.

Three independent reviewer runs were synthesized. The statistical analysis dimension was not applicable because no inferential statistics were used. The copyedit pass flagged 30 issues (8 major). The verification components found 2 dead links and a tortured phrase indicating possible text generation.

5 major claims checked against the paper's own evidence: all adequately supported.

1 integrity concern flagged (0 high).

Reporting

Eight Dimensions3 pass · 1 warn · 1 fail

The paper reviews prior sign language recognition methods, noting their reliance on costly instruments, plain backgrounds, and high computational power. It then presents a clear rationale for using a multi-headed CNN with image and hand landmark fusion to address these issues. The limitations of prior work, such as dependence on single-channel CNN and plain backgrounds, are explicitly discussed and mitigated by the proposed approach.

Sub-criteria

prior work citedADEQUATE

premise rationaleADEQUATE

limitations addressedADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 95% conf

The paper clearly establishes the scientific foundation by citing prior work, providing a logical rationale for the fusion approach, and addressing limitations of existing methods.

Evidence

paraphrase[Introduction, paragraphs 1-3 and Literature Review section]

“Many methods are available that use different things/tools like images (2D, 3D), sensor data (hand globe, Kinect sensor, neuromorphic sensor), videos, etc. All things are considered due to the fact that the captured images are excessively noisy.”

direct quote[Introduction, paragraph 4]

“In this work, we have applied a fusion of traditional image processing with extracted hand landmarks and trained on a multi-headed CNN so that it could complement each other’s weights on the concatenation layer.”

direct quote[Literature Review, last paragraph]

PASSReviewer 2· 90% conf

The paper establishes a clear scientific foundation by citing relevant prior work, providing a logical rationale for the study, and addressing limitations of previous methods.

Evidence

direct quote[Introduction, paragraph 2]

“State-of-the-art techniques centered after utilizing deep learning models to improve good accuracy and less execution time.”

paraphrase[Introduction, paragraph 4]

“The authors state that their main objective is to achieve a better detection rate without relying on a traditional single-channel CNN, and that this method has been proven to work well with less computational power on medical image datasets.”

direct quote[Introduction, paragraph 1]

PASSReviewer 3· 90% conf

The study clearly establishes the need for cost-effective sign language recognition and reviews prior work, highlighting limitations such as reliance on costly instruments or simple backgrounds.

Evidence

direct quote[Abstract]

“However, such drawbacks can be easily overcome by employing artificial intelligence-based techniques. Since, in this modern era of advanced mobile technology, using a camera to take video or images is much easier, this study demonstrates a cost-effective technique to detect American Sign Language (ASL) using an image dataset.”

direct quote[Introduction, paragraph 3]

“The available online datasets are already processed or taken in a lab environment where it becomes easy for recent advanced AI models to train and evaluate, causing prone to errors in real-life applications with different kinds of noises. Accordingly, it is a basic need to make a model that can deal with noisy images and also be able to deliver positive results.”

direct quote[Literature review, paragraph 5]

“The literature summarizes that most of the models used in this application either depend on a single variable or require high computational power. Also, their dataset choice for training and validating the model is in plain background, which is easier to detect. Our main aim is to show how to reduce the computational power for training and the dependency of model training on one layer.”

For a computational study, the applicable sub-criteria are power_analysis (dataset selection), inclusion_exclusion (data curation), controls (baseline/ablations), outlier_handling, and independent_replication. The dataset is described with a 70-30 train-test split, but no formal power analysis or sample size justification is provided. Inclusion/exclusion criteria are clear (24 letters, excluding j and z). The paper presents ablation results comparing the combined model to image-only and hand-landmark-only models, serving as adequate controls. Outlier handling is not mentioned, and independent replication is not reported. Thus 2 of 5 applicable sub-criteria are adequate, yielding a warn.

Sub-criteria

controlsADEQUATE

power analysisNOT REPORTED

blinding levelsN/A

outlier handlingNOT REPORTED

randomization unitN/A

inclusion exclusionADEQUATE

randomization methodN/A

replicate distinctionN/A

independent replicationNOT REPORTED

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 90% conf

The study design is appropriate for a computational benchmark: dataset selection is justified, inclusion/exclusion of classes is specified, and baseline models serve as controls.

Evidence

direct quote[Materials and methods, Dataset description]

“We intended to avoid that single color background and use a complex background with many users’ hand images to increase the detection complexity.”

direct quote[Abstract and Materials and methods]

“with 24 letters (except j and z as they contain motion)”

direct quote[Result analysis, paragraph 4]

“Before combining image and hand landmark inputs, we tested both individually and acquired a test accuracy of 96.29% for the image and 98.42% for hand landmarks.”

WARNReviewer 2· 80% conf

The study design has several gaps: no power analysis, no outlier handling, and no independent replication; however, inclusion/exclusion criteria and baseline controls are adequately reported.

Evidence

paraphrase[Methods, Dataset description]

“We have used the 'ASL Finger Spelling' dataset, which has images of different sizes, orientations, and complex backgrounds of over 500 images per sign (24 sign total) of 4 users.”

paraphrase[Results, Table 5]

“Table 5 compares the combined model with image-only and hand-landmark-only models in terms of parameters, execution time, memory, and accuracy.”

absence[Methods, Training and testing]

No power analysis or outlier handling is reported.

WARNReviewer 3· 70% conf

The study describes the dataset, pre-processing steps, and model architecture in detail, but lacks a formal power analysis and clear distinction between biological and technical replicates.

Evidence

direct quote[Materials and methods, Dataset description, paragraph 1]

“That’s why we have used the “ASL Finger Spelling” dataset , which has images of different sizes, orientations, and complex backgrounds of over 500 images per sign (24 sign total) of 4 users (non-native to sign language).”

direct quote[Abstract]

“To avoid the overfitting problem, data augmentation and dynamic learning rate reduction have been used.”

direct quote[Training and testing, paragraph 1]

“For training, we have used 46,023 images, and for testing, 19,725 images.”

The study uses a pre-existing image dataset for sign language recognition. While the dataset contains images of human hands, the study itself does not involve direct manipulation or collection of biological variables, making this dimension not applicable.

Sub-criteria

demographicsN/A

sex reportedN/A

sex justifiedN/A

age weight healthN/A

housing conditionsN/A

species strain sourceN/A

Per reviewer (3) — the votes behind the verdict

N/AReviewer 1· 95% conf

The study uses a public image dataset of human hands but does not report any biological variables such as sex, age, or health status of the subjects. As a computational study on image data, these variables are not applicable.

Evidence

direct quote[Materials and methods, Dataset description]

“over 500 images per sign (24 sign total) of 4 users (non-native to sign language).”

N/AReviewer 2· 95% conf

The study uses only image data and does not involve biological subjects or variables, so this dimension is not applicable.

N/AReviewer 3· 100% conf

This study is computational, using an image dataset, so biological variables are not directly applicable.

Evidence

direct quote[Abstract]

“this study demonstrates a cost-effective technique to detect American Sign Language (ASL) using an image dataset.”

The paper explicitly states the use of a publicly available dataset from Kaggle. This dataset consists of images and does not involve the collection of new human or animal subject data, nor does it use non-public identifiable data.

Sub-criteria

iacuc statementN/A

informed consentN/A

irb ethics statementN/A

regulatory complianceN/A

Per reviewer (3) — the votes behind the verdict

N/AReviewer 1· 95% conf

The study uses a publicly available, de-identified dataset from Kaggle. No new human subject data was collected, so ethical approvals are not required.

Evidence

paraphrase[Data availability]

“The dataset used in this paper (ASL Fingerspelling Images (RGB & Depth)) is publicly available at Kaggle on this URL: https://www.kaggle.com/datasets/mrgeislinger/asl-rgb-depth-fingerspelling-spelling-it-out.”

N/AReviewer 2· 95% conf

The study uses a publicly available dataset from Kaggle with no new collection of human subject data, so ethics approval is not required.

Evidence

direct quote[Data availability]

N/AReviewer 3· 100% conf

The study uses a publicly available, de-identified image dataset and does not involve new human or animal subject data collection, making ethical approvals not applicable.

Evidence

direct quote[Data availability]

For a computational study, the applicable sub-criterion is software_tools_identified. The paper mentions Google's hand landmark model, ImageDataGenerator, Adam optimizer, and the Keras/TensorFlow framework, but does not provide version numbers, URLs, or RRIDs. No custom code is deposited in a public repository. Thus, the single applicable sub-criterion is reported_but_inadequate, leading to a fail.

Sub-criteria

mycoplasma testingN/A

reagents identifiedN/A

organisms identifiedN/A

antibodies identifiedN/A

cell line authenticationN/A

software tools identifiedINADEQUATE

Per reviewer (3) — the votes behind the verdict

WARNReviewer 1· 70% conf

The only applicable key resource is software tools, which are identified but not with version numbers or permanent identifiers, making the reporting inadequate.

Evidence

direct quote[Materials and methods, Hand landmark detection]

“Google’s hand landmark model has an input channel of RGB and an image size of (224 × 224 × 3).”

direct quote[Materials and methods, Model building]

“This model has been compiled with Adam optimizer and MSE loss for 50 epochs.”

direct quote[Materials and methods, Training and testing]

“Image Data Generator did image augmentation with 10° rotation, 0.1 zoom range, 0.1 widths and height shift range, and horizontal flip.”

FAILReviewer 2· 85% conf

The only applicable resource is software tools, which are named but not identified with versions or RRIDs, and custom code is not shared, resulting in a fail.

Evidence

paraphrase[Methods, Hand landmark detection]

“Google's hand landmark model has an input channel of RGB and an image size of (224 × 224 × 3).”

absence[Data availability]

No code repository is provided.

PASSReviewer 3· 90% conf

The study adequately identifies the dataset, pre-trained models, and software used, including specific versions and sources where applicable.

Evidence

direct quote[Data availability]

direct quote[Hand landmark detection, paragraph 1]

“Google’s hand landmark model has an input channel of RGB and an image size of (224 × 224 × 3).”

direct quote[Model building, paragraph 2]

“This model has been compiled with Adam optimizer and MSE loss for 50 epochs.”

The paper reports test accuracy, per-class precision, recall, and F1 scores from a confusion matrix. No hypothesis tests, p-values, or effect sizes with confidence intervals are reported. The data presentation is adequate (confusion matrix, per-class metrics). Since no inferential statistics are used, the entire dimension is not applicable.

Sub-criteria

tests namedN/A

exact p valuesN/A

effect sizes ciN/A

data presentationADEQUATE

software identifiedN/A

assumptions verifiedN/A

mathematical plausibilityN/A

Per reviewer (3) — the votes behind the verdict

N/AReviewer 1· 90% conf

The study reports classification metrics (accuracy, precision, recall, F1) but does not perform any inferential statistics (p-values, confidence intervals). Therefore, no statistical analysis criteria apply.

Evidence

paraphrase[Result analysis, Table 3 and Figure 7]

“The Precision, Recall, and F1 score for 24 classes are shown in Table 3.”

N/AReviewer 2· 90% conf

The paper reports classification metrics (accuracy, precision, recall, F1) and a confusion matrix, but no inferential statistical tests are performed, so this dimension is not applicable.

PASSReviewer 3· 85% conf

The study clearly names the metrics used (precision, recall, F1 score, accuracy) and provides detailed results in tables and figures, with per-class support counts.

Evidence

direct quote[Training and testing, paragraph 2]

“For further evaluation, we have calculated the precision, recall, and F1 score of the proposed multi-headed CNN model, which shows excellent performance.”

direct quote[Training and testing, paragraph 2]

“To compute these values, we first calculated the confusion matrix (shown in Fig. ).”

paraphrase[Table 3]

“Table 3 shows the Precision, Recall, and F1 score for 24 classes, along with the support for each class.”

The paper includes a dedicated 'Data availability' section that explicitly states the dataset is publicly available and provides a direct URL to the Kaggle repository. No custom code is explicitly mentioned as being shared, but the dataset is the primary data resource.

Sub-criteria

code sharingNOT REPORTED

accession numbersN/A

repository depositADEQUATE

data availability statementADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 85% conf

The data availability statement is concrete, naming the public repository with a URL. Code is not shared, but the dataset is fully accessible.

Evidence

paraphrase[Data availability]

absence[Data availability and elsewhere]

No code sharing statement found.

WARNReviewer 2· 85% conf

The dataset URL is provided, but custom code is not shared, leading to a warn.

Evidence

direct quote[Data availability]

absence[Data availability]

No code is shared.

PASSReviewer 3· 95% conf

The study provides a clear data availability statement with a direct link to the public Kaggle dataset.

Evidence

direct quote[Data availability]

Methods are described in sufficient detail for replication, including dataset preprocessing, model architecture, training parameters, and augmentation. All outcomes (accuracy, precision, recall, F1) are reported. Limitations are discussed in the conclusion. Conclusions are proportional to the evidence, with no overclaiming. Funding source is stated, but a conflicts-of-interest statement is absent. Five of six applicable sub-criteria are adequate, yielding a pass.

Sub-criteria

funding coiINADEQUATE

trial registrationN/A

reporting guidelineN/A

methods completenessADEQUATE

all outcomes reportedADEQUATE

limitations discussedADEQUATE

conclusions proportionalADEQUATE

Per reviewer (3) — the votes behind the verdict

PASSReviewer 1· 85% conf

The paper provides detailed methods, reports all outcomes, discusses limitations, and states conclusions proportionally. A conflicts-of-interest statement is missing, but funding is reported.

Evidence

direct quote[Materials and methods, Model building]

“Two-dimensional Convolutional layers with filter size 50, 25, kernel (3, 3) with Relu, strides 1; MaxPooling 2D with pool size (2, 2), batch normalization, and Dropout layer has been used in the hand landmarks training side.”

direct quote[Conclusion]

“Some limitations found in the proposed method compared with the literature. Some methods might work with low image dataset numbers, but as we use the simple CNN model, this method requires a good number of images for training.”

direct quote[Funding]

“Funding was provided by the American University of the Middle East, Egaila, Kuwait.”

PASSReviewer 2· 85% conf

The methods are detailed, outcomes are fully reported, limitations are discussed, and conclusions are proportional; however, a conflict of interest statement is missing, which is a minor gap.

Evidence

direct quote[Methods, Model building]

direct quote[Conclusion]

“Some methods might work with low image dataset numbers, but as we use the simple CNN model, this method requires a good number of images for training.”

paraphrase[Funding]

“The funding section states the source but does not include a conflict of interest declaration.”

WARNReviewer 3· 75% conf

The methods are generally well-described for replication, and limitations are discussed, but specific reporting guidelines are not referenced, and the statistical software used is not identified.

Evidence

direct quote[Conclusion, paragraph 3]

“There are some limitations found in the proposed method compared with the literature. Some methods might work with low image dataset numbers, but as we use the simple CNN model, this method requires a good number of images for training. Also, the proposed method depends on the hand landmark extraction model. Other hand landmark model can cause different results.”

direct quote[Funding]

“Funding was provided by the American University of the Middle East, Egaila, Kuwait.”

paraphrase[Materials and methods]

“The 'Materials and methods' section provides detailed descriptions of the dataset, image pre-processing steps, and the multi-headed CNN model architecture, including layer types, filter sizes, and activation functions.”

30 copyedit issues flagged (8 major): mostly grammar, clarity, typo.

Checked 32 references by DOI: 26 verified — 6 no DOI (shown, not verified).

Reproducibility

2 data/code links checked; 0 live, 2 dead.

Αlpha¹ Rigor Action ItemsFinal pass before journal submission

As a published paper, an informed reader should weigh the following: (1) the data availability link is dead, undermining reproducibility; (2) software versions are not reported, making the computational environment irreproducible; (3) the text contains a tortured phrase ('convolutional neural organization') that is a fingerprint of paraphrasing tools, warranting scrutiny of the manuscript's originality. These issues would warrant an erratum to fix the URL and add a COI statement, and ideally a public code repository for re-analysis.

1.
HIGHdata code
Fix the dead data availability link (currently https://www.kaggle.com/datasets/mrgeislinger/asl-rgb-depth-fingerspelling-spelling-it-out) or provide an alternative permanent repository (e.g., Zenodo) with a DOI.
The reproducibility verification found both links dead; readers cannot access the dataset without a working link.
2.
HIGHother
Investigate and replace the tortured phrase 'convolutional neural organization' (which appears in the text as a paraphrase of 'convolutional neural network') with the correct term, and review the entire manuscript for similar unnatural paraphrases that may indicate text generation or paper mill activity.
Such phrases are a known fingerprint of paraphrasing tools and undermine the integrity of the manuscript.
3.
HIGHcopyedit
Rewrite the Abstract and Introduction for clarity and correct grammar, including fixing the misuse of 'quiet' for 'muteness', 'handicapped' for 'disabled', and 'implacable' (should be 'applicable') in the Conclusion.
These errors significantly affect readability and professional presentation; a reader may question the rigor of the work if the writing is poor.
4.
HIGHreporting
Add a conflict of interest statement in the Declarations section (e.g., 'The authors declare no competing interests.') and ensure it is compliant with the journal's requirements.
This is a standard reporting requirement; its absence is a minor but noticeable gap.
5.
HIGHdata code
Deposit the custom code (multi-headed CNN model, training, and evaluation scripts) in a public repository (e.g., GitHub, Zenodo) and cite it in the Data Availability section.
Code sharing is essential for reproducibility of computational studies; its absence is a major gap.
6.
MEDIUMrigor
Add version numbers and source URLs for all software tools used (e.g., Python, TensorFlow/Keras, OpenCV, Google MediaPipe, scikit-learn) in the Methods section.
Without version information, the computational environment is not reproducible.
7.
MEDIUMrigor
Add a statement on whether any power analysis or sample size justification was performed, or justify why none is needed for this benchmark study.
A formal power analysis is not always required for computational benchmarks, but its absence should be acknowledged.
8.
MEDIUMreporting
Complete Table 1 by filling in the missing depth and intensity data for sessions B-E, or clarify that these values are only available for session A.
An incomplete table undermines the transparency of the dataset description.
9.
MEDIUMreporting
Add a statement on outlier handling (e.g., 'No outliers were excluded from the dataset') in the Methods section.
Readers need to know whether any data were excluded and why.
10.
MEDIUMcopyedit
Fix the awkward phrasing in the Abstract: 'it requires many costly instruments' should be 'they require many costly instruments' (subject-verb agreement with 'studies').
Grammatical errors reduce the professionalism of the paper.
11.
LOWreporting
Add a section on data curation detailing any inclusion/exclusion criteria for individual images (e.g., blurry images, unclear hands) to improve study design transparency.
This would strengthen the description of dataset preparation.
12.
LOWstatistics
If the exact test set accuracy is 98.98%, report the number of correctly classified images out of the total test set (e.g., '19,525 out of 19,725') to allow verification.
This provides a simple check for the reported accuracy.

Full text (34,528 chars)Analyzed: Aug 5, 2026

Start the discussion on JRNLClub

Community discussion on jrnlclub.com

Read alongside