REVIEW 2 major objections 1 minor 4 references
Unsupervised Machine Learning for Osteoporosis Diagnosis Using Singh Index Clustering on Hip Radiographs
T0 review · 2 major / 1 minor · reviewed 2026-05-23 · grok-4.3
Pith's one-line read Unsupervised clustering on custom CNN features from unlabeled hip X-rays groups images into Singh Index grades for osteoporosis diagnosis.
desk verdict The paper runs a custom CNN on 838 unlabeled hip X-rays then clusters into six groups hoping to match Singh Index grades, but only two clusters score well on internal metrics and nothing ties the groups to actual clinical labels or DEXA values. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The custom convolutional neural network architecture for feature extraction from radiographs, which supplies the input representations for unsupervised clustering into SI grades without any labeled training examples.
What would settle it
Expert radiologists assign Singh Index grades to the same 838 images and the unsupervised clusters match those labels for fewer than three grades overall.
Extended reading notes
Core claim
A custom convolutional neural network extracts features from an unlabeled dataset of 838 hip X-ray images that enable various clustering algorithms to categorize the images into six Singh Index grade clusters, with the custom architecture showing superior cluster homogeneity and heterogeneity compared to established models, although only two clusters display high Silhouette Scores indicative of promising classification.
Load-bearing premise
Radiographic image features alone contain enough information to recover all six Singh Index grades via unsupervised clustering, without requiring labeled examples or additional clinical variables such as patient age.
Editorial extensions
If this is right
- Dataset imbalance limits reliable separation into all six SI grades.
- Image quality directly affects the homogeneity of the resulting clusters.
- Adding patient clinical data alongside the radiographs improves classification accuracy.
- Image pre-processing and semi-supervised learning methods can reduce reliance on large labeled sets.
Reading between the lines
- The approach could function as an initial binary screen (high versus low risk) rather than a full six-grade diagnostic tool if only two clusters separate cleanly.
- Applying the same pipeline to radiographs from other populations or scanners would test whether the two reliable clusters generalize beyond the current Indian adult cohort.
- Integrating the clustering output with existing clinical workflows might reduce the number of cases requiring full expert SI review.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript claims to automate Singh Index (SI) grading for osteoporosis diagnosis from hip radiographs via unsupervised ML. A custom CNN extracts features from an unlabeled set of 838 images; multiple clustering algorithms are then applied with k=6 to recover the six SI grades. The custom CNN is reported to outperform established models on homogeneity/heterogeneity metrics, yet only two of the six clusters achieve high silhouette scores. The authors note dataset imbalance and recommend augmenting the pipeline with clinical variables and image pre-processing.
Significance. A validated unsupervised pipeline could reduce reliance on DXA for initial screening in low-resource settings. The current empirical results, however, demonstrate only partial recovery of the target grades, so the work primarily illustrates the difficulty of the unsupervised task and the value of hybrid clinical-image approaches rather than delivering a ready automation method.
major comments (2)
- [Abstract] Abstract: the explicit statement that 'only two clusters with high Silhouette Scores' achieved promising classification directly undermines the central claim that the pipeline categorizes images into six SI grades. This limitation is load-bearing because the automation goal requires reliable separation across all grades, not a subset.
- [Results] Results (and Methods): no external validation against expert SI annotations or DXA BMD is performed on the unlabeled 838-image dataset. Internal metrics alone cannot establish that the clusters recover trabecular patterns rather than acquisition artifacts or demographic confounders, which is required to support the diagnostic claim.
minor comments (1)
- [Abstract] Abstract: the phrase 'promising classification' for the two high-silhouette clusters should be replaced by a quantitative statement of the actual silhouette values and the fraction of the dataset they cover.
Simulated Author's Rebuttal
We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below, with plans for revisions where the manuscript requires clarification or adjustment to align claims with results.
read point-by-point responses
-
Referee: [Abstract] Abstract: the explicit statement that 'only two clusters with high Silhouette Scores' achieved promising classification directly undermines the central claim that the pipeline categorizes images into six SI grades. This limitation is load-bearing because the automation goal requires reliable separation across all grades, not a subset.
Authors: We agree that the abstract phrasing should be revised for precision. The manuscript already reports that only two clusters achieved high silhouette scores, indicating that full separation into six grades was not attained. We will update the abstract to explicitly note this partial outcome, temper the automation claim to reflect demonstrated potential only for select grades, and highlight the challenges posed by dataset imbalance and image quality. This ensures the stated objectives match the empirical findings without overstating the pipeline's current capabilities. revision: yes
-
Referee: [Results] Results (and Methods): no external validation against expert SI annotations or DXA BMD is performed on the unlabeled 838-image dataset. Internal metrics alone cannot establish that the clusters recover trabecular patterns rather than acquisition artifacts or demographic confounders, which is required to support the diagnostic claim.
Authors: The referee accurately notes the absence of external validation, which stems directly from the unlabeled nature of the 838-image dataset. Internal metrics (silhouette scores, homogeneity, heterogeneity) were used to evaluate cluster quality, but these cannot confirm recovery of true SI trabecular patterns versus confounders. We will expand the discussion section to explicitly state this limitation, recommend incorporation of labeled data or clinical variables for validation in future extensions, and clarify that the current work illustrates the unsupervised task's difficulties rather than claiming a validated diagnostic tool. revision: partial
- Absence of external validation on the unlabeled dataset, as no expert SI annotations or DXA BMD values are available to confirm clusters recover trabecular patterns rather than artifacts or confounders.
Circularity Check
No circularity; empirical unsupervised clustering on unlabeled radiographs with internal metrics only.
full rationale
The paper reports an empirical application of a custom CNN for feature extraction followed by standard clustering algorithms (k=6) on an unlabeled set of 838 hip X-ray images. Evaluation relies solely on internal metrics such as Silhouette Score, homogeneity, and heterogeneity. No equations, derivations, fitted parameters renamed as predictions, or self-citations appear in the load-bearing steps. The central results are direct outputs of applying off-the-shelf unsupervised methods to the given data; they do not reduce to the inputs by construction. This is the most common honest finding for purely empirical ML papers without theoretical claims.
Assumptions & free parameters
free parameters (2)
- Number of clusters =
6
- Custom CNN architecture and training hyperparameters
assumptions (1)
- domain assumption Convolutional neural networks extract features from hip radiographs that reflect trabecular patterns used in the Singh Index.
Cite this review
Pith. "Pith review of Unsupervised Machine Learning for Osteoporosis Diagnosis Using Singh Index Clustering on Hip Radiographs." pith.science (2026). https://pith.science/paper/2411.15253
@misc{pith2026241115253,
author = {Pith},
title = {Pith review of: Unsupervised Machine Learning for Osteoporosis Diagnosis Using Singh Index Clustering on Hip Radiographs},
year = {2026},
howpublished = {\url{https://pith.science/paper/2411.15253}},
note = {Machine review of arXiv:2411.15253}
}
read the original abstract
Osteoporosis, a prevalent condition among the aging population worldwide, is characterized by diminished bone mass and altered bone structure, increasing susceptibility to fractures. It poses a significant and growing global public health challenge over the next decade. Diagnosis typically involves Dual-energy X-ray absorptiometry to measure bone mineral density, yet its mass screening utility is limited. The Singh Index (SI) provides a straightforward, semi-quantitative means of osteoporosis diagnosis through plain hip radiographs, assessing trabecular patterns in the proximal femur. Although cost-effective and accessible, manual SI calculation is time-intensive and requires expertise. This study aims to automate SI identification from radiographs using machine learning algorithms. An unlabelled dataset of 838 hip X-ray images from Indian adults aged 20-70 was utilized. A custom convolutional neural network architecture was developed for feature extraction, demonstrating superior performance in cluster homogeneity and heterogeneity compared to established models. Various clustering algorithms categorized images into six SI grade clusters, with comparative analysis revealing only two clusters with high Silhouette Scores for promising classification. Further scrutiny highlighted dataset imbalance and emphasized the importance of image quality and additional clinical data availability. The study suggests augmenting X-ray images with patient clinical data and reference images, alongside image pre-processing techniques, to enhance diagnostic accuracy. Additionally, exploring semi-supervised and self-supervised learning methods may mitigate labelling challenges associated with large datasets.
Reference graph
Works this paper leans on
-
[1]
Proposed Model for Feature Extraction CNNs like DenseNet, InceptionV3, and EfficientNet are among the first used to extract features from the X-ray images. When K-Means clustering is used, these networks, however, produce low Silhouette Scores, which point to less-than-ideal outcomes. As a result, a unique CNN architecture is developed, which consists of ...
work page 2021
-
[2]
Conclusion The present study aimed to conduct an extensive performance analysis on the application of ML models for predicting SI using pelvic X-ray images. Rather than utilizing DXA images, which are routinely employed to diagnose osteoporosis by calculating BMD, this investigation opted for X-ray images. The objective was to develop an automated system ...
-
[3]
Convolutional Neural Network for Second Metacarpal Radiographic Osteoporosis Screening
Fathima, Nazia & Tamilselvi, R. & Beham, Dr.M.Parisa. (2019). Estimation of t-score and BMD values from X-ray images for detection of osteoporosis. 220-224. 10.1145/3309074.3309110. [19] Sato, Y., Yasuhiko Takegami, Takamune Asamoto, Ono, Y., Tsugeno Hidetoshi, Goto, R., Kitamura, A. and Honda, S. Artificial intelligence improves the accuracy of residents...
-
[4]
Rethinking the Inception Architecture for Computer Vision,
Nissinen T, Suoranta S, Saavalainen T, Sund R, Hurskainen O, Rikkonen T, Kröger H, Lähivaara T, Väänänen SP. Detecting pathological features and predicting fracture risk from dual-energy X-ray absorptiometry images using deep learning. Bone Rep. 2021;14:101070. doi: 10.1016/j.bonr.2021.101070. [47] Deepak Saini, Trilok Chand, Devendra K. Chouhan, Mahesh P...
Reviewed May 23, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.