XAI-DermNet: A Dual-Modality Deep Learning and Explainable

AI Fusion Framework for Transparent Skin Lesion Diagnosis

from Dermoscopic Images

Sukkirtha K.1,* Anbuchelian S.1 John Prakash A.1

1 Ramanujan Computing Centre, Anna University, Chennai, Tamil Nadu, India

Received: January 17, 2026 Revised: February 15 2026 Accepted: March 19, 2026 ⋆ Corresponding author

ABSTRACT

The advancement of trustworthy diagnostic tools in dermatological automation is hindered by the limited transparency

of current deep learning systems, which function as opaque models and impede clinical acceptance. This research

presents a novel intelligent framework for skin lesion analysis that unites deep learning methodologies with

explainable artificial intelligence (XAI) principles to address this interpretability deficit. The proposed approach

utilizes a transfer-learned ResNet50 architecture for robust image classification, coupled with Local Interpretable

Model-agnostic Explanations (LIME) to furnish clear, visual justifications for the model’s outputs. Performance

assessment on the HAM10000 benchmark yielded a classification accuracy of 94.3%, with a validation accuracy

of 91.8%. Concurrently, the LIME framework effectively identified and visualized diagnostically critical features

in the lesion images, thereby elucidating the model’s reasoning process for medical practitioners. These findings

confirm that augmenting high-performance deep learning with post-hoc explanatory techniques yields a credible

and understandable diagnostic instrument, thereby promoting clinician trust and facilitating data-informed medical

judgments. Subsequent developments will prioritize scalable cloud implementation, interoperability with healthcare

information systems, extension to underrepresented lesion categories, and rigorous evaluation in diverse clinical

environments.

Keywords: Skin lesion diagnosis Deep learning ResNet50 Explainable artificial intelligence LIME

1. INTRODUCTION

Skin diseases are among the most prevalent health

conditions worldwide, affecting individuals of all ages

and skin types. Early and accurate diagnosis plays a

crucial role in providing effective treatment and

preventing the progression of severe dermatological

conditions such as melanoma and other ma-lignant skin

lesions. Traditionally, dermatologists rely on visual

inspection and dermoscopic examination to identify skin

abnormalities. However, this diagnostic approach can

sometimes be subjective and depends heavily on the

expertise and experience of the clinician. In addition,

manual examination can be time-consuming when large

numbers of patients require screening. As a result, there

is increasing interest in developing automated diagnostic

systems that can assist der-matologists in detecting and

classifying skin diseases more efficiently and

consistently. Advances in medical imaging and

computational techniques have created opportunities to

improve diagnostic accuracy through intelligent

computer-aided systems. [5]