PaperlayerView exact revision on arXiv

Source-linked AI summary

Detail in Context: A Dual-Scale Machine Learning Framework for Mycosis Fungoides Detection

Mohamed Hazem, Tarek Waleed, Omar Khaled, Nada Omar, Mahmoud Raslan, Marwa Mohamed Fawzy, Aya Fahim, Rania M. Mogawer, Ahmed Mourad, Kariman Mansour, Muhammad Rushdi

arXiv:2609.38560v1cs.CV

TL;DR

Early MF can resemble benign inflammatory dermatoses, and MF remains underexplored in AI-assisted dermatologic diagnosis. The paper combines dual-magnification image models with clinical-feature classifiers; image late fusion and clinical random forest results are reported for detection, with clinical models also evaluated for MF staging.

Abstract

from arXiv · show

Mycosis fungoides (MF) is a rare form of cutaneous T-cell lymphoma that is often misdiagnosed in early stages due to its visual similarity to benign inflammatory dermatoses. Early and accurate diagnosis is critical for improving patient outcomes. In this paper, we propose a comprehensive diagnostic framework for automated MF detection that combines dual- scale histopathological image analysis with deep learning. To distinguish MF from other lymphoproliferative skin conditions, the proposed approach leverages a late-fusion ensemble of dual- magnification (10x and 20x) convolutional neural networks (CNNs), complemented by a random forest classifier trained on 16 clinical features. Experimental results on an expanded dataset of 6,267 images (4,306 MF; 1,961 Non-MF) across 463 patients demonstrate that strong detection performance is obtained by prioritizing higher-resolution cytological details (20x) within broader architectural context (10x). The image-based late-fusion model achieves an accuracy of 83.58% and a sensitivity of 89.13%, while the clinical random forest model achieves an accuracy of 96.6% and sensitivity of 93.8%, highlighting the po- tential of this multimodal framework as a robust clinical decision support system in dermatology. This framework addresses two distinct clinical objectives: an image-based dual-scale pipeline optimized for the early diagnostic screening of MF versus non- MF dermatoses, and a complementary clinical metadata model designed for the subsequent staging of confirmed MF cases (patch/plaque versus tumor)

I. INTRODUCTION

Early MF can resemble benign inflammatory skin disorders, while MF remains underexplored in AI-assisted dermatologic diagnosis. The paper proposes dual-scale image analysis alongside clinical metadata models for distinct diagnostic tasks.

I. INTRODUCTION

  • A separate model trained on 16 clinical features complements image analysis for disease progression and staging in established MF cases.The paper compares random forests, XGBoost, and logistic regression.Source: §I · p. 1

II. RELATED WORK

  • AI research has largely focused on common skin cancers, while MF-focused work remains limited and MF is often underrepresented in studies of lymphoproliferative skin diseases.The paper also contrasts its fixed-magnification conventional microscopy approach with WSI-based methods, addressing scanner constraints in resource-limited settings.Source: §I · p. 1, §II · pp. 1–2

I. INTRODUCTION

  • The framework uses 10× and 20× histopathology CNNs, with weighted late fusion prioritizing cellular detail while retaining architectural context.The contributions also include patient-level mean aggregation of patch probabilities.Source: §I · p. 1

III. MATERIALS AND METHODS

The study uses retrospective histopathology images and associated metadata, divided at the patient level for model development and evaluation. Its dataset covers multiple MF-mimicking disorders and two magnifications.

A. Dataset

  • The dataset comprises 6,257 images from 463 patients, including 2,146 images at 10× and 4,121 at 20× magnification.Images were captured from regions of interest using a conventional microscope.Source: A. Dataset · p. 2
  • The non-MF cohort includes a heterogeneous range of lymphoproliferative skin disorders that mimic MF clinically and histologically.Table I is identified as presenting patient and image distributions across diagnostic categories and magnifications.Source: A. Dataset · p. 2, Table I · p. 2
  • Patients were partitioned into 329 training, 67 validation, and 67 test patients.The images were further subdivided into patches; 20× images yielded 23,062 training, 4,728 validation, and 4,320 test patches, while 10× images yielded 73,824, 14,804, and 12,848, respectively.Source: A. Dataset · p. 2
  • Figure 1 illustrates different foreground ratios in patches from a sample TIF image.The figure caption identifies example patches rather than reporting a quantitative result.Source: Figure 1 · p. 2

B. Preprocessing

Preprocessing extracts patches at each magnification while preserving the 20× images’ higher optical resolution. The modeling setup uses separate branches to capture architectural and cellular features.

B. Preprocessing

  • 10× images were tiled into 512×512 patches with a 256-pixel stride; 20× images used 1024×1024 patches with a 512-pixel stride before downsampling to 512×512.The 20× optical capture preserves superior resolution and cytological detail despite digital resizing.Source: B. Preprocessing · p. 2
  • HSV saturation filtering discarded patches with foreground pixel ratios below 28.5%, while training augmentations varied spatial orientation and color.RGB mean-intensity thresholding was insufficient under staining and illumination inconsistencies; the figure shows sample foreground ratios.Source: B. Preprocessing · p. 2

1) Dual-Magnification Architecture:

  • Independent TF-EfficientNet-B3 branches analyze 10× architectural distortion and 20× cellular atypia across five diagnostic classes.The models have approximately 10.3 million parameters; Table II is identified as summarizing magnification-specific configurations and hyperparameters.Source: 1) Dual-Magnification Architecture · p. 3, Table II · p. 3
  • Both the 10× architectural and 20× cytological models were trained using the AdamW optimizer, with a static learning rate of 2 × 10−4 for each model.Source: 2) Training Strategy and Augmentation · p. 3

2) Training Strategy and Augmentation:

Training uses class-weighted loss and augmentation, with validation F2-score guiding checkpoint selection. Patient-level predictions are formed by averaging patch probabilities.

2) Training Strategy and Augmentation:

  • Class weights were set inversely to training-class representation, and training used spatial and color augmentation to address imbalance and improve generalization.Augmentations included flips, rotations, and color jitter; the passage also notes mixed precision and gradient accumulation for memory use and batch-gradient stability.Source: 2) Training Strategy and Augmentation · p. 3

3) Metric Selection and Checkpointing:

  • Both models selected checkpoints using validation-set macro F2-score, chosen to emphasize sensitivity and penalize false negatives.The passage describes the F2-score as weighting recall twice as heavily as precision.Source: 3) Metric Selection and Checkpointing · p. 3

D. Patient-Level Prediction and Decision Thresholding

  • For each patient, the model averages the predicted probabilities across that patient’s patches to produce a patient-level score.This aggregation is presented as the step for obtaining clinically meaningful patient-level predictions.Source: D. Patient-Level Prediction and… · p. 3

1) Threshold Optimization via Youden’s J Statistic:

The 10× and 20× models required different decision thresholds, and late fusion combined their patient-level probabilities with greater weight on 20× cellular detail. The selected fusion balance retained 20× emphasis while incorporating 10× architectural context.

1) Threshold Optimization via Youden’s J Statistic:

  • At a 0.5 threshold, the 20× model’s MF sensitivity was 43.48%; Youden’s J selected 0.8351, restoring 89.13% sensitivity and 83.58% accuracy without degrading non-MF detection.The distinct operating points necessitated piecewise-linear probability calibration for subsequent late fusion.Source: 1) Threshold Optimization via Youden’s… · p. 3

1) Late Fusion Approach:

  • Late fusion aggregates independent model probabilities, letting 10× capture architectural disarray and 20× capture cellular atypia without feature-space contamination.Unlike early fusion, it combines predictions after each network specializes at its magnification.Source: 1) Late Fusion Approach · p. 3

2) Weighted Score Aggregation and Clinical Justification:

3) Optimization of the Fusion Weight:

The clinical classifier uses selected metadata features in a sequential pipeline: diagnosis of MF versus Non-MF, followed by staging among patients classified as MF. Random forests were selected based on cross-validation performance.

F. Clinical Features Classifier

  • Sixteen features with p < 0.05 were selected from 22 clinical variables, while Nodule was retained despite statistical insignificance for binary MF diagnosis because of its staging value.The selected features cover demographics, disease history, anatomic distribution, and lesion morphology.Source: F. Clinical Features Classifier · p. 4, Table III · p. 4
  • The sequential pipeline diagnoses MF versus Non-MF in all 499 records, then predicts patch–plaque versus tumor stage for the 353 MF records using the same feature set.The 16 features were selected through univariate testing, with Nodule retained for its staging value.Source: 1) Task 1 — Diagnosis (MF vs. Non-MF) · p. 4, F. Clinical Features Classifier · p. 4
  • Random forests, XGBoost, and logistic regression were compared using 5-fold stratified cross-validation, and random forests were selected for both tasks based on overall performance.The pipeline uses median imputation for missing continuous values, label encoding for categorical variables, and adjusted positive-class weights.Source: 1) Task 1 — Diagnosis (MF vs. Non-MF) · p. 4

IV. RESULTS

On a held-out test set of 67 patients, the late-fusion model achieved 83.58% accuracy and an F2-score of 0.8874. Fig. 3 compares patch-level and patient-level metrics at both magnifications.

A. Deep Learning Image Analysis Results

  • The held-out test set comprised 67 patients, with 12,848 patches at 10× and 4,320 patches at 20× magnification.The results section identifies these as the evaluation data for the proposed framework.Source: A. Deep Learning Image Analysis Results · p. 4
  • Fig. 3 compares patch-level metrics in dark blue with patient-level metrics in dark orange across both magnifications.The figure caption specifies the two aggregation levels and their colors.Source: Figure 3 · p. 5
  • 83.58% accuracy and an F2-score of 0.8874 were achieved using the late-fusion model.Table IV summarizes quantitative performance on a test set of 67 patients.Source: A. Deep Learning Image Analysis Results · p. 5, Table IV · p. 5

B. Patient-Level Aggregation Benefits

Patient-level aggregation improves predictions across metrics, and the 20× model’s cytological detail supports sensitive detection. Clinical models perform strongly for diagnosis and staging, with SHAP analysis highlighting influential diagnostic features.

B. Patient-Level Aggregation Benefits

C. Sensitivity and Clinical Relevance

  • The 20× model achieved F2-score 0.8874, compared with 0.8222 for the 10× model.The reported rationale emphasizes cellular details including hyperchromatic, cerebriform nuclei.Source: C. Sensitivity and Clinical Relevance · p. 5

D. Clinical Feature Classification Results

V. DISCUSSION

The 20× model outperformed the 10× model on F2-score, while the late-fusion approach combines cellular detail with lower-magnification architectural context. For staging, the Nodule feature dominates feature importance.

A. Cellular Detail within Architectural Context

  • The 20× model achieved F2=0.8874, compared with F2=0.8222 for the 10× model.The paper associates higher magnification with better resolution of cytological features such as hyperchromatic, irregular cerebriform nuclei.Source: A. Cellular Detail within… · p. 6
  • For MF staging, Nodule is the single most important predictor, despite lacking statistical significance in the diagnosis task.The paper identifies nodules and tumors as hallmarks of advanced-stage MF.Source: Figure 6 · p. 6
  • Late fusion assigns 80% decision weight to the 20× cytological model and 20% to the 10× architectural model.The lower magnification captures architectural patterns such as epidermotropism and Pautrier’s microabscesses.Source: A. Cellular Detail within… · p. 6

B. The Necessity of Whole-Slide Context

Mean aggregation integrates patch-level signals to support patient-level diagnosis. The authors note that image-only diagnosis simplifies clinical practice, while their fixed-magnification approach avoids whole-slide digitization and reports competitive performance.

B. The Necessity of Whole-Slide Context

  • Mean aggregation integrates signals across patches, reflecting MF features that are subtly spread through tissue rather than concentrated in isolated hotspots.The paper likens this to a pathologist scanning the whole slide before diagnosis.Source: B. The Necessity of Whole-Slide Context · p. 6

C. Benchmarking and Clinical Reality

D. Comparison with State-of-the-Art

  • 83.58% accuracy, 89.13% sensitivity, and ROC-AUC 0.797 were reported using standard fixed-magnification microscopy without whole-slide digitization.The authors contrast this with recent systems using whole-slide images or specialized equipment, which they say can limit deployment in resource-constrained settings.Source: D. Comparison with State-of-the-Art · pp. 6–7

VI. CONCLUSION

The conclusion reports strong image-based and clinical-model results, with 20× cytological detail prioritized in late fusion. Future work proposes direct multimodal integration and stain normalization.

VI. CONCLUSION

  • With 80% fusion weight on 20×, the late-fusion ensemble achieves 89.13% sensitivity and 83.58% accuracy; the clinical random forest achieves 93.8% sensitivity and 96.6% accuracy.The protocol combines 20× cytological detail with 10× architectural context through patient-level mean aggregation.Source: §VI · p. 7

VII. FUTURE WORK

  • Future work proposes cross-attention architectures to integrate histopathological features directly with clinical tabular data, moving beyond independent models.Source: §VII · p. 7
  • Future iterations will incorporate stain normalization, such as Macenko or Vahadane, to mitigate confounding effects from H&E color variation across laboratory settings.Source: §VII · p. 7

Related papers

Loading 2609.38560v1…