PaperlayerView exact revision on arXiv

Source-linked AI summary

An AI-Based Multi-Stage Approach for Androgenetic Alopecia Assessment from Low-Magnification Scalp Images

Mahmoud Raslan, Nada Omar, Omar Khaled, Tarek Waleed, Mohamed Hazem, Rania Mounir, Solwan Elsamanoudy, Ahmed Mourad, Noura Adel, Muhammad Rushdi

arXiv:2610.02421v1cs.CV

TL;DR

Manual trichoscopic interpretation of AGA features remains subjective, motivating quantitative clinical support. The paper integrates follicular-unit detection, shaft counting and width measurement, regional aggregation, and a rule layer; its detector and count classifier report strong task-level results, while the framework remains decision support for dermatologists.

  • Problem

    Trichoscopic assessment can identify AGA-relevant features, but manual interpretation remains subjective and varies between observers.

    Source: §I · p. 1

  • Method

    The framework integrates follicular-unit localization, ordinal shaft counting, calibrated width measurement, regional aggregation, and an interpretable rule-based layer for dermatologist review.

    Source: E. Multi-Region Clinical Decision… · p. 5, §VII · p. 7

  • Results

    YOLOv8m achieved mAP@0.5=0.920 and recall=0.860; EfficientNet-B5+S achieved 87% expert-box count accuracy (macro F1=0.85).

    Source: §VI · p. 7

  • Takeaways & Limitations

    The framework is quantitative clinical decision support for dermatologist interpretation, not an autonomous disease classifier.

    Source: E. Multi-Region Clinical Decision… · p. 5, §VI · p. 7

    The study has not completed patient-level comparisons between rule outputs and all 243 clinician diagnoses, so diagnostic accuracy, sensitivity, specificity, and kappa are not reported.

    Source: §VI · p. 7

Abstract

from arXiv · show

Androgenetic alopecia (AGA) is characterized by patterned follicular miniaturization, increased single-hair follicular units, and altered hair-shaft diameter. We present an automated quantitative scalp-analysis and clinical decision-support framework combining FU localization, ordinal visible-shaft counting, calibrated shaft-width estimation, regional aggregation, and an interpretable rule layer. The clinical cohort comprised 243 patients (127 AGA, 116 non-AGA), while the computer-vision experiments used 160 expert-annotated patients, 2,400 trichoscopic images, and approximately 158,000 FU annotations. Under patientdisjoint evaluation, YOLOv8m achieved test mAP@0.5=0.920 and recall=0.860; EfficientNet-B5 with a support-map channel achieved 87.0% expert-box count accuracy (macro F1=0.85). A separate 500-image set was processed end-to-end with detector-generated boxes, yielding MAE of 6.56 for follicle detection and 16.59 for follicle classification relative to human-expert annotations. The system is intended to assist, rather than replace, dermatologist interpretation.

I. INTRODUCTION

AGA assessment relies on measuring patterned hair changes, but trichoscopic interpretation remains manual and variable. This work integrates quantitative scalp measurements into an interpretable decision-support workflow for clinicians.

I. INTRODUCTION

  • Reliable AGA severity assessment matters for treatment planning and monitoring, while manual trichoscopic interpretation introduces subjectivity and inter-observer variability.Trichoscopy can reveal follicular density, hair-diameter variation, and miniaturized hairs.Source: §I · p. 1
  • The framework combines follicular-unit localization, ordinal visible-shaft counting, calibrated shaft-width measurement, and regional aggregation in a rule-based clinical support workflow.Frontal measurements are interpreted relative to an occipital reference for the AGA use case.Source: §I · p. 1, Figure 1 · p. 2
  • Small, densely packed follicular units, overlapping shafts, and dandruff complicate quantitative analysis of scalp images.The study focuses on these content-related challenges under standardized image acquisition.Source: §I · p. 1
  • Overlapping tiles enlarge follicular units in detector inputs, while a support map emphasizes filamentous structures and root-anchored tracing measures shafts from detected units.These strategies address small targets and shaft overlap without relying on unattributed global skeleton branches.Source: §I · p. 1
  • The contribution is an integrated, interpretable measurement framework—not a standalone learned AGA-versus-non-AGA classifier—and its outputs are intended to support dermatologists.The authors note that extension to other scalp conditions would require appropriately redefined clinical criteria.Source: §I · p. 1

II. RELATED WORK

Prior scalp-image analysis spans classical image processing and deep learning, while small targets, overlapping shafts, and complex texture remain challenging. The study integrates existing methods and evaluates them using a standardized, patient-disjoint imaging cohort distinct from its broader clinical cohort.

II. RELATED WORK

  • Earlier scalp-image methods range from classical phototrichograms and handcrafted features to deep learning, with small targets, dense shaft overlap, and complex texture remaining challenges.Slicing-aided inference has been used to improve small-object recall in high-resolution images.Source: §II · p. 2
  • The work integrates localization, ordinal counting, calibrated width measurement, and multi-region aggregation rather than proposing a novel backbone; broader cross-family benchmarking remains future work.Internal comparisons focus on several YOLO and EfficientNet configurations.Source: §II · p. 2

A. Clinical Cohort and Image Acquisition

  • The clinical cohort comprised 243 patients: 127 with AGA and 116 without AGA, including FPHL and MPHL cases and TE and clinically normal comparison cases.The AGA group included 102 FPHL and 25 MPHL patients; the non-AGA group included 63 TE and 53 clinically normal patients.Source: A. Clinical Cohort and Image Acquisition · p. 2
  • The expert-annotated imaging dataset contained 2,400 images from 160 patients, acquired at five scalp landmarks using a standardized handheld dermatoscope and smartphone setup.Three images were obtained at each landmark per patient.Source: A. Clinical Cohort and Image Acquisition · p. 2
  • A patient-disjoint 70/15/15 split assigned 112 patients to training and 24 each to validation and testing, keeping each patient's images and derived crops together.Acquisition conditions were standardized, and dandruff was identified as a naturally occurring visual obstacle represented in the dataset.Source: A. Clinical Cohort and Image Acquisition · p. 2
  • Per-region follicular-unit annotation counts are excluded as quantitative evidence pending an additional clinical audit and planned dataset-release preparation.De-identified summaries and an annotated image subset may be shared subject to institutional, consent, and data-sharing approval.Source: A. Clinical Cohort and Image Acquisition · p. 2

B. Annotation and Derived Task Datasets

Experts annotated visible follicular units with bounding boxes and ordinal shaft-count labels. The study derived classifier crops from those annotations and reserved a separate image set to examine end-to-end processing.

B. Annotation and Derived Task Datasets

  • Experts delineated approximately 158,000 follicular units and labeled each as one, two, or three-plus visible shafts.Each visible follicular unit received an axis-aligned bounding box.Source: B. Annotation and Derived Task Datasets · p. 3
  • Classifier crops expanded expert boxes by 10 pixels per side; crops smaller than 16 pixels in either dimension were discarded, yielding 126,626 crops.Crop assignment followed the parent patient's split.Source: B. Annotation and Derived Task Datasets · p. 3
  • A separate 500-image set propagated detector-generated boxes through the classifier and morphometric modules to examine the automated chain and regional summaries.Without independent expert labels for every generated crop, the set was not used for a second crop-level classification-accuracy estimate.Source: B. Annotation and Derived Task Datasets · p. 3

IV. METHODOLOGY

The methodology passes quantitative measurements through four computer-vision stages and a clinical aggregation stage. Preprocessing isolates the scalp region and uses overlapping tiles to support follicular-unit detection.

A. Overall Pipeline

  • The pipeline comprises follicular-unit localization, ordinal shaft counting, shaft-width estimation and hair-type categorization, regional aggregation, and a final rule-based AGA-support summary.The stages pass measurements forward to an interpretable clinical output.Source: A. Overall Pipeline · p. 3

1) Task objective and input:

  • Follicular-unit localization is a single-class small-object detection task intended to provide spatial coverage for downstream count and width analysis..Source: 1) Task objective and input · p. 3

2) Preprocessing:

  • Preprocessing extracts and expands a scalp-bearing region, then divides it into overlapping 640 × 640 tiles with 30% overlap to enlarge follicular units in detector inputs.CLAHE improves local luminance contrast, and background-only tiles are capped.Source: 2) Preprocessing · p. 3

3) Model choice and architecture:

The detector uses a slicing-aided strategy for small follicular units, while the classifier combines RGB crops with a support map that emphasizes shaft-like structures. Count accuracy is evaluated using expert-annotated boxes, so it does not represent end-to-end accuracy.

3) Model choice and architecture:

  • YOLOv8m is the selected detector, with all follicular-unit subtypes collapsed into one class because the ordinal classifier handles count subtype.The model was selected for patient-level localization performance and moderate complexity.Source: 3) Model choice and architecture · p. 3

4) Output and implementation:

1) Task objective and evaluation distinction:

  • The 87% count-classification accuracy is measured on expert-annotated boxes and should not be interpreted as end-to-end system accuracy.Automated operation instead generates crops from YOLO predictions.Source: 1) Task objective and evaluation… · p. 4

2) Support-map generation:

  • The support map combines normalized darkness, local contrast, edge, black-hat, and ridge cues into an auxiliary channel for shaft-like structures.The same fixed generator is used for classification and width estimation; the map supplements rather than replaces RGB appearance.Source: 2) Support-map generation · p. 4

3) Architecture and model choice:

The study prioritizes EfficientNet for ordinal visible-shaft classification and passes its predicted count downstream as a structural constraint on shaft-path selection.

3) Architecture and model choice:

  • EfficientNet variants and support-map-augmented versions were explored, while exhaustive transformer and hybrid-backbone comparisons were left for future work.EfficientNet was prioritized for accuracy-per-parameter efficiency, fine-grained classification suitability, and compatibility with the ordinal CORAL head.Source: 3) Architecture and model choice · p. 4

4) Output passed downstream:

  • The predicted one-, two-, or three-plus-shaft count constrains how many proximal shaft candidates the width-estimation stage retains.This constraint is used when several plausible paths emerge from one follicular unit.Source: 4) Output passed downstream · p. 4

1) Calibration and task objective:

The width-estimation workflow calibrates image coordinates, compares a global branch-based baseline with root-anchored local tracing, and measures shaft diameter near the proximal shaft.

1) Calibration and task objective:

  • The valid imaging field is perspective-warped into an upright crop, and pixel distances are converted to micrometers using the spacer slit dimensions.The slit is specified as 16 × 6.25 mm.Source: 1) Calibration and task objective · p. 4

2) Global branch-based baseline:

  • The global baseline fuses hair masks from the full image and local follicular-unit crops to recover strands that may be suppressed in full-image processing.Candidate branches are ranked by root-endpoint distance, outward extension, and local support, then width is estimated from proximal profiles.Source: 2) Global branch-based baseline · pp. 4–5

3) Root-anchored support-map method:

  • The root-anchored method extracts a padded local crop, tests support-map thresholds, and accepts a mask when components connect to the root region and extend beyond the follicular-unit box.Candidate paths are traced outward and filtered for pass-through, duplication, sharp bends, and unstable support.Source: 3) Root-anchored support-map method · p. 5
  • Diameters are measured perpendicular to the selected centerline, primarily around 250 µm from the root, and retained only when profiles are consistent and locally supported.Measurements affected by crossings, overlap, or dandruff near the path are rejected; diameters are categorized as vellus, intermediate, or terminal.Source: 3) Root-anchored support-map method · p. 5

E. Multi-Region Clinical Decision Support

The rule-based layer organizes regional trichoscopic measurements for dermatologist review, using frontal measurements relative to an occipital reference for the AGA use case.

E. Multi-Region Clinical Decision Support

  • The rule-based layer compares frontal measurements with an occipital reference because AGA is a patterned miniaturization disorder,,,.It is intended to organize measured features for dermatologist review, not serve as a learned AGA-versus-non-AGA classifier.Source: E. Multi-Region Clinical Decision… · p. 5
  • The regional summary includes the percentage of single-hair follicular units, hair distribution width, and hair-shaft diameter variation.The supplied table caption identifies clinician-facing AGA-support categories, but does not state their values.Source: E. Multi-Region Clinical Decision… · p. 5, Table III · p. 5
  • The listed AGA criteria include frontal hair-shaft diameter variation above 20% and a frontal single-hair-unit percentage more than twice the occipital percentage.The cited rationale links increased vellus burden and frontal predominance of single-hair units with miniaturization and reduced multi-hair-unit density,, –.Source: E. Multi-Region Clinical Decision… · p. 5
  • Regional counts, density, hair distribution width, diameter variation, and shaft-width distributions are displayed for dermatologist staging or grading in clinical context.The outputs are not intended as autonomous diagnoses.Source: E. Multi-Region Clinical Decision… · p. 5

V. EXPERIMENTAL RESULTS

The final detector result was achieved by YOLOv8m, while support-map EfficientNet-B5 led the controlled expert-box count-classification evaluation.

A. Follicular-Unit Detection

  • YOLOv8m achieved held-out test mAP@0.5 of 0.920 and recall of 0.860 after focused optimization under the patient-disjoint protocol.Table IV documents earlier YOLO-family screening, not the subsequent focused optimization that produced the reported final result.Source: A. Follicular-Unit Detection · p. 5

B. Follicular-Count Classification

  • Support-map EfficientNet-B5 achieved 87.0% accuracy, macro precision 0.86, macro recall 0.85, and macro F1 0.85 on controlled crop-level evaluation.Table V compares EfficientNet classifiers under expert-verified FU localization; the reported 87% does not measure end-to-end counting from detector-generated boxes.Source: B. Follicular-Count Classification · pp. 5–6, Table V · p. 6
  • Adding the support-map channel raised accuracy for B2 from 0.827 to 0.835, B3 from 0.839 to 0.842, and B5 from 0.860 to 0.870.EfficientNet-B5+S provided the strongest overall result.Source: B. Follicular-Count Classification · p. 6
  • The classifier benchmark measures visible-shaft counting only when an expert has already localized a true FU, excluding detector errors and localization offsets.Expert annotations do not provide count labels for every YOLO-generated proposal.Source: B. Follicular-Count Classification · p. 6

C. End-to-End Processing on the Additional 500-Image Set

The additional 500-image set was processed through the full automated sequence, with task-output MAEs reported against human-expert annotations. The width experiment found a runtime advantage for root-anchored measurement, but its unequal agreement subsets do not support a controlled accuracy comparison.

C. End-to-End Processing on the Additional 500-Image Set

  • On the additional 500-image set, the automated pipeline achieved MAE of 6.56 for follicle detection and 16.59 for follicle classification against human-expert annotations.The experiment used detector-generated boxes and evaluated system-level operation rather than controlled expert-box classification accuracy.Source: C. End-to-End Processing on the… · p. 6
  • Figure 5 color-codes FU boxes by predicted ordinal count and renders measured shaft segments by diameter category.It shows representative end-to-end output from the automated processing sequence.Source: Figure 5 · p. 6

D. Hair-Width Estimation

  • The root-anchored method produced 158 valid measurements from 160 reference strands at 12.0 s/image, while global branch-based processing took 142.8 s/image.The global method's agreement analysis used only a 15-strand subset, so the study does not present the agreement statistics as a controlled head-to-head accuracy comparison.Source: D. Hair-Width Estimation · p. 6

E. Clinical Output and Visualization

The framework presents clinically interpretable regional measurements as dermatologist-facing support, while reported module performance is qualified by validation and benchmarking limits. Rule outputs do not replace clinician diagnosis or staging.

E. Clinical Output and Visualization

  • Regional measurements and frontal-to-occipital comparisons map to support categories, but final diagnosis and formal staging remain the dermatologist’s responsibility.Automated rule outputs have not been paired with all patient-level diagnoses, so diagnostic accuracy, sensitivity, specificity, and kappa are not reported.Source: E. Clinical Output and Visualization · p. 6

VI. DISCUSSION

  • YOLOv8m reached mAP@0.5=0.920 and recall=0.860; EfficientNet-B5+S achieved 87% expert-box count accuracy (macro F1=0.85), and the 500-image predicted-box experiment yielded system-level MAE of 6.56 for detection and 16.59 for classification.The discussion describes patient-disjoint evaluation and focused YOLOv8 development.Source: §VI · pp. 6–7
  • Increased one-hair FU proportion reflects loss of multi-hair density, while vellus burden, diameter diversity, and frontal-to-occipital comparison capture miniaturization and patterned AGA distribution.Acquisition was standardized; the main content-related challenges were small or dense FUs, shaft overlap or crossing, and dandruff.Source: §VI · p. 7
  • Width validation remains preliminary, the two width methods use unequal reference subsets, the FotoFinder comparison is small, and cross-family architecture benchmarking is limited.The patient-level comparison between rule outputs and all 243 clinician diagnoses is also incomplete.Source: §VI · p. 7

VII. CONCLUSION AND FUTURE WORK

  • Future work prioritizes larger FotoFinder and expert width validation, dermatologist-reviewed proposals and clinical outputs, broader backbone comparisons, multi-center evaluation, and condition-specific extension to other scalp disorders.The framework is intended as quantitative clinical decision support rather than an autonomous disease classifier.Source: §VII · p. 7, §VI · p. 7
Loading 2610.02421v1…