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

cs.CVSep 29, 2026

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

Authors: Mohamed Hazem, Tarek Waleed, Omar Khaled, Nada Omar, Mahmoud Raslan, Marwa Mohamed Fawzy, Aya Fahim, Rania M. Mogawer, +3 more

Organizations: Faculty of Engineering, Cairo University, Giza, Egypt · Dermatology Department, Faculty of Medicine, Cairo University, Cairo, Egypt

Abstract

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)

Figures & tables

Diagnostic Category	No. of Patients	No. of Images
Mycosis Fungoides (MF)	311	4,306
Non-MF Cohort (Total)	152	1,961
B-cell lymphoma	18	358
PLEVA / PLC	108	1,268
T-cell dyscrasia	15	180
Pseudolymphoma	11	155

TABLE I: Distribution of patients and images across diagnostic categories and magnifications.

Fig. 1: Different foreground ratios for patches of a sample TIF image

Parameter	10 $\times$ Model	20 $\times$ Model
Base Architecture	TF-EfficientNet-B3	TF-EfficientNet-B3
Input Resolution	$512\times 512$	$512\times 512$
Dropout Rate	0.40	0.40
Inference Batch Size	8	2
Validation Metric	F 2 -Score	F 2 -Score

TABLE II: Model Configuration and Hyperparameters

Fig. 2: Optimization of multimodal late fusion weights. The graph illustrates the impact of varying the 10 $\times$ model’s weight ( $w$ ) on overall Accuracy and MF-specific F2-Score. The clinical optimum was identified at $w=0.2$ , demonstrating that a fusion ratio heavily weighted toward 20 $\times$ cytological features (80%) augmented by 10 $\times$ architectural context (20%) optimally balances high accuracy with the strict sensitivity required for MF screening.

Clinical Feature	Data Type	Statistical Test	$p$ -value
Selected Features (Significant)
Macules	Categorical	Chi-Squared	$<0.001$
Biopsy 1 Morphology	Categorical	Chi-Squared	$<0.001$
Papules	Categorical	Chi-Squared	$<0.001$
Age	Continuous	Mann-Whitney U	$<0.001$
Color	Categorical	Chi-Squared	$<0.001$

TABLE III: Univariate statistical analysis of clinical features ranked by

p

-value. Features with

p<0.05

were selected; Nodule was retained despite

p>0.05

for its staging value (see text).

	10 $\times$ Model	20 $\times$ Model	Late Fusion ( $\alpha=0.20$ )
Accuracy	0.8060	0.8358	0.8358
F 2 -Score	0.8222	0.8874	0.8874
Sensitivity	0.8043	0.8913	0.8913
Specificity	0.8043	0.8913	0.8913
ROC-AUC	0.8685	0.8509	0.8778

TABLE IV: Quantitative Performance Summary (Test Set, N=67 Patients)

Fig. 3: A side-by-side comparison of patch-level (dark blue) vs. patient-level (dark orange) metrics across both magnifications.

Metric	XGBoost	Random Forest	Logistic Reg.
Accuracy	0.956	0.966	0.950
Precision	0.908	0.948	0.915
Sensitivity	0.945	0.938	0.918
F 2 -Score	0.937	0.939	0.916
ROC-AUC	0.994	0.995	0.985

TABLE V: Task 1: MF vs. Non-MF Diagnosis (5-Fold Stratified CV)

Fig. 4: SHAP feature importance for Task 1 (MF vs. Non-MF). Features ranked by the mean absolute SHAP value.

Fig. 5: SHAP beeswarm plot for Task 1. Each dot represents a single patient; color indicates feature value (red = high, blue = low), and horizontal position shows the SHAP value (impact on prediction).

Metric	XGBoost	Random Forest	Logistic Reg.
Accuracy	0.972	0.977	0.952
Precision	0.828	0.870	0.700
Sensitivity	0.860	0.860	0.527
F 2 -Score	0.843	0.855	0.532
ROC-AUC	0.990	0.987	0.968

TABLE VI: Task 2: MF Staging Results — Patch–Plaque vs. Tumor (5-Fold Stratified CV)

Fig. 6: Feature importance for Task 2 (Stage of MF). The Nodule feature dominates, confirming its near-deterministic role in identifying the tumor stage.

Study	Data Modality	Clinical Task	Reported Metric	Deployment Requirement
Doeleman et al. (2024) [ 5 ]	WSI (H&E)	Early MF vs. Benign	Mean AUC: 0.827	High (Digital WSI Scanner)
Zhao et al. (2026) [ 9 ]	WSI + Clinical Data	MF vs. Benign Inflammatory	Macro-AUC: $>$ 0.85	High (Digital WSI Scanner)
Liu et al. (2025) [ 10 ]	Dermoscopy + Clinical Images	Early MF vs. Inflammatory	Accuracy: 82.9%*	Medium (Dermoscopy equipment)
Ghosh et al. (2026) [ 11 ]	Non-Linear Optical Microscopy	Detection of Epidermotropism	Qualitative	Very High (NLOM infrastructure)
Ours	Fixed-Magnification (10 $\times$ , 20 $\times$ ) H&E	MF vs. Non-MF	Accuracy: 83.58%	Low (Standard Optical Microscope)
*Performance reported for the Dermatologist + AI assisted group.

TABLE VII: Comparison of Recent AI Frameworks for Mycosis Fungoides (MF) Diagnosis

Explore similar work

JI-ADF: Joint-Individual Learning with Adaptive Decision Fusion for Multimodal Skin Lesion Classification
Apr 30, 2026Phan Nguyen, Dat Cao, Hien Kha +4Skin Lesion Classification Melanoma
1KAIST · 2Taipei Medical University · 3Yale University
A Comparative Evaluation of Pre-trained Convolutional Neural Networks for Melanoma Detection
Sep 12, 2026Wagner Moreno Schmitz, Marco Antonio de Castro Barbosa, Thiago Magalhães Amaral +2Skin Lesion Classification Melanoma
Universidade Federal do Vale do São Francisco (Univasf), Petrolina, Pernambuco, Brasil · Universidade Tecnológica Federal do Paraná (UTFPR), Pato Branco, Paraná, Brasil
Multimodal Skin Lesion Classification with Swin Transformer and Clinical Metadata Fusion
Aug 4, 2026Nethmi Pathirana, Isuru Munasinghe, Dileeka AlwisSkin Lesion Classification Lesion Segmentation
School of Computing, Informatics Institute of Technology, Colombo, Sri Lanka · Department of Electronic and Telecommunication Engineering, University of Moratuwa, Katubedda, Sri Lanka

Apr 30, 2026, cs.CVJ/K move · Enter open · S save

JI-ADF: Joint-Individual Learning with Adaptive Decision Fusion for Multimodal Skin Lesion Classification

Phan Nguyen, Dat Cao, Hien Kha, Hien Chu+3

1KAIST · 2Taipei Medical University · 3Yale University

Skin Lesion Classification Melanoma Adaptive Feature Fusion

Open paper AskPDF arXiv

Skin lesion classification is essential for early dermatological diagnosis, yet many existing computer-aided systems rely primarily on dermoscopic images and underutilize the multimodal evidence routinely available in clinical practice. To address this gap, we propose \textbf{JI-ADF}, a trimodal deep learning framework that integrates dermoscopic images, clinical photographs, and structured patient metadata for clinically grounded skin lesion classification. The proposed architecture combines joint multimodal representation learning with modality-specific auxiliary supervision and an adaptive decision fusion mechanism that dynamically calibrates modality contributions on a per-sample basis. To enhance cross-modal reasoning while preserving modality-specific evidence, we further introduce a multimodal fusion attention (MMFA) module. We evaluate JI-ADF on the large-scale MILK10k benchmark, which reflects real-world clinical acquisition conditions and severe class imbalance. The proposed method demonstrates strong and well-balanced performance across lesion categories, improving sensitivity and Dice score while maintaining high specificity and good calibration. Extensive analyses, including modality ablation, calibration evaluation, and Grad-CAM visualization, further confirm the robustness and clinically meaningful behavior of the model. These results indicate that JI-ADF provides a reliable and practical foundation for multimodal skin lesion classification in real-world clinical settings.

Abstract

Explore similar work

JI-ADF: Joint-Individual Learning with Adaptive Decision Fusion for Multimodal Skin Lesion Classification

A Comparative Evaluation of Pre-trained Convolutional Neural Networks for Melanoma Detection

Multimodal Skin Lesion Classification with Swin Transformer and Clinical Metadata Fusion