Masked Privileged-Information Distillation for Multimodal Skin Lesion Classification Under Missing Clinical Metadata

Anirban Barua; Md Mahir Abrar Khan; Ayman Iktidar; Md. Sajjatul Islam

cs.CVOct 2, 2026

Masked Privileged-Information Distillation for Multimodal Skin Lesion Classification Under Missing Clinical Metadata

Authors: Anirban Barua, Md Mahir Abrar Khan, Ayman Iktidar, Md. Sajjatul Islam

Organizations: Department of Computer Science and Engineering Chittagong Independent University (CIU), Chattogram, Bangladesh

Abstract

Multimodal skin lesion classification combines clinical images with patient metadata to improve diagnostic accuracy. However, complete metadata available during training may be only partially accessible at deployment, and resource-constrained settings additionally require computational efficiency. We address these challenges with a privileged-information distillation framework in which a multimodal teacher trained on complete metadata supervises a 9.2x smaller student trained with randomly masked clinical fields. Clinical fields are masked as whole groups at a per-sample rate drawn from U(0,1), so one training run covers the full metadata availability range. Fusion is residual, with metadata added as a gated correction to an unconditional image base. On the PAD-UFES-20 dataset, distillation under masked training improves balanced accuracy over cross-entropy training at every availability level, by an average of +4.7 points versus +1.8 points without masking. The masked student loses only 7.8 balanced-accuracy points as metadata decreases from complete to absent, compared with 36.6 points for the same student trained on complete metadata, highlighting the role of masked training in graceful degradation beyond distillation alone. Grad-CAM visualizations further show that the masked student's attention generally remains lesion-centered as metadata is withdrawn. The resulting compact model targets point-of-care settings, where clinical metadata is often incomplete.

Figures & tables

Fig. 1: Overview of the proposed privileged-information distillation framework

Class	Train	Validation	Test
ACK	523	106	101
BCC	590	126	129
MEL	38	6	8
NEV	168	37	39
SCC	131	31	30
SEK	158	41	36

TABLE I: Data distribution in train, validation, and test sets

Model	BACC	Macro AUC
Metadata only	0.628 $\pm$ 0.039	0.895 $\pm$ 0.007
Image only	0.673 $\pm$ 0.036	0.931 $\pm$ 0.003
MetaBlock [ 4 ]	0.758 $\pm$ 0.030	0.948 $\pm$ 0.008
Concatenation	0.790 $\pm$ 0.033	0.963 $\pm$ 0.005
Residual Gated Fusion	0.821 $\pm$ 0.006	0.965 $\pm$ 0.001

TABLE II: Performance of ConvNeXt-Base models at full metadata availability

Fig. 2: Balanced accuracy across metadata availability for distilled and cross-entropy students, under masked (left) and unmasked (right) training

Availability	KD (Proposed)	CE
0.0	0.706 $\pm$ 0.016	0.675 $\pm$ 0.007
0.2	0.730 $\pm$ 0.011	0.688 $\pm$ 0.012
0.4	0.743 $\pm$ 0.009	0.695 $\pm$ 0.011
0.6	0.763 $\pm$ 0.009	0.709 $\pm$ 0.016
0.8	0.775 $\pm$ 0.010	0.720 $\pm$ 0.021
1.0	0.784 $\pm$ 0.019	0.733 $\pm$ 0.028

TABLE III: Balanced accuracy of the distilled (KD) and cross-entropy (CE) students under masked training

Availability	KD (Proposed)	CE
0.0	0.435 $\pm$ 0.015	0.487 $\pm$ 0.051
0.2	0.571 $\pm$ 0.017	0.565 $\pm$ 0.028
0.4	0.675 $\pm$ 0.020	0.640 $\pm$ 0.017
0.6	0.743 $\pm$ 0.017	0.695 $\pm$ 0.014
0.8	0.775 $\pm$ 0.007	0.746 $\pm$ 0.014
1.0	0.801 $\pm$ 0.015	0.771 $\pm$ 0.018

TABLE IV: Balanced accuracy of the distilled (KD) and cross-entropy (CE) students under unmasked training

Metric	Teacher	Student (Proposed)
Parameters (M)	89.0	9.7
Size (MB)	356.2	38.9
GMACs	15.4	1.4
CPU latency (ms)	217.8	35.4
GPU latency (ms)	13.6	5.2
BACC ( $r{=}0$ )	0.452	0.706

TABLE V: Teacher and distilled student on size, compute, and accuracy.

Fig. 3: Grad-CAM attributions for the masked and unmasked students across metadata availability, for three lesions.

Explore similar work

Privileged Lesion-Context Relational Distillation for Mask-Free Skin Lesion Classification
Jul 21, 2026Abu Mukaddim Rahi, Md Mithun Hossain, Md Zulficar Hasan Joy +2Skin Lesion Classification Medical Image Classification
Department of IT in Artificial Intelligence, Victoria University, Footscray, VIC, Australia · BUBT Resarch Graduate School, Bangladesh University of Business and Technology, Dhaka, Bangladesh · Department of Artificial Intelligence, Daffodil Institute of Information Technology, Dhaka, Bangladesh +2
DerMAE: Improving skin lesion classification through conditioned latent diffusion and MAE distillation
Feb 23, 2026Francisco Filho, Kelvin Cunha, Fábio Papais +6Skin Lesion Classification
Centro de Inform´atica, Universidade Federal de Pernambuco, Brazil
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 Medical Diagnosis
1KAIST · 2Taipei Medical University · 3Yale University

Jul 21, 2026J/K move · Enter open · S save

Privileged Lesion-Context Relational Distillation for Mask-Free Skin Lesion Classification

Abu Mukaddim Rahi, Md Mithun Hossain, Md Zulficar Hasan Joy, M. F. Mridha+1

Department of IT in Artificial Intelligence, Victoria University, Footscray, VIC, Australia · BUBT Resarch Graduate School, Bangladesh University of Business and Technology, Dhaka, Bangladesh · Department of Artificial Intelligence, Daffodil Institute of Information Technology, Dhaka, Bangladesh+2

Skin Lesion Classification Medical Image Classification Relational Knowledge Distillation Learning Using Privileged Information

Open paper AskPDF arXiv

Accurate skin lesion classification can benefit from lesion segmentation masks, but requiring masks or an auxiliary segmentation model during inference reduces clinical practicality and increases computational complexity. This work introduces Privileged Lesion-Context Relational Distillation (PLCRD), a teacher-student framework that exploits lesion masks exclusively during training while preserving image-only inference. The privileged teacher jointly analyzes the original dermoscopic image and its mask-guided lesion region to learn lesion-specific and contextual diagnostic representations. An image-only student is then trained through complementary knowledge-transfer mechanisms that convey the teacher's diagnostic distribution, lesion-focused attention, inter-lesion relational geometry, and lesion-context structure. PLCRD decomposes deep representations into lesion and contextual embeddings and transfers their relational organization through inter-lesion similarity alignment, lesion-context affinity matching, separation regularization, and class-aware relational learning. This formulation avoids direct feature matching between heterogeneous teacher and student architectures and enables the student to internalize mask-informed diagnostic structure without accessing masks at deployment. The framework was evaluated on HAM10000 using lesion-disjoint data partitioning and externally validated on ISIC 2018 without retraining. PLCRD achieved a lesion-level macro-F1 of 0.773 +/- 0.018, balanced accuracy of 0.764 +/- 0.023, and macro-AUROC of 0.976 +/- 0.002 on HAM10000, together with a macro-F1 of 0.732 +/- 0.008 on ISIC 2018. The results indicate that privileged lesion annotations can be transformed into transferable relational knowledge, yielding a practical and interpretable approach to mask-free skin lesion classification.

Abstract

Explore similar work

Privileged Lesion-Context Relational Distillation for Mask-Free Skin Lesion Classification

DerMAE: Improving skin lesion classification through conditioned latent diffusion and MAE distillation

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