Bridging the Black Box: An Interpretable Deep Learning Framework for Multi-Class Dermatological Screening

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autori principali: Mr. Aniket Kumar, Mr. Pawan Kumar
Natura: Recurso digital
Pubblicazione: Zenodo 2026
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866901544558919680
author Mr. Aniket Kumar
Mr. Pawan Kumar
author_facet Mr. Aniket Kumar
Mr. Pawan Kumar
contents <p class="MsoNormal"><strong><span>Abstract</span></strong></p> <p class="MsoNormal"><span>Skin cancer remains a major clinical concern, and timely assessment of suspicious lesions can improve dermatological decision-making. This paper presents DermaSense AI, a deep learning framework for automated multi-class skin lesion classification on HAM10000. The proposed system combines a two-phase transfer-learning strategy with Xception augmented by a Convolutional</span></p> <p class="MsoNormal"><span>Block Attention Module (CBAM) and DenseNet201. A weighted soft-voting ensemble achieves a macro Area Under the Curve (AUC) of 95.92% and an overall accuracy of 80.24% on the held-out test set. To support qualitative interpretation, Gradientweighted Class Activation Mapping (Grad-CAM) is used to visualize model attention. In addition, a dedicated melanoma-screening analysis attains a Negative Predictive Value (NPV) of 97.1% at the selected operating threshold, highlighting the potential of the framework as a decision-support aid rather than a standalone diagnostic system.</span></p> <p class="MsoNormal"><strong><span>Keywords:</span></strong><span> Skin Lesion Classification, Deep Learning, CBAM Attention, Ensemble Learning, Grad-CAM, HAM10000, Clinical Triage</span></p> <p class="MsoNormal"><span> </span></p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_20259265
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Bridging the Black Box: An Interpretable Deep Learning Framework for Multi-Class Dermatological Screening
Mr. Aniket Kumar
Mr. Pawan Kumar
<p class="MsoNormal"><strong><span>Abstract</span></strong></p> <p class="MsoNormal"><span>Skin cancer remains a major clinical concern, and timely assessment of suspicious lesions can improve dermatological decision-making. This paper presents DermaSense AI, a deep learning framework for automated multi-class skin lesion classification on HAM10000. The proposed system combines a two-phase transfer-learning strategy with Xception augmented by a Convolutional</span></p> <p class="MsoNormal"><span>Block Attention Module (CBAM) and DenseNet201. A weighted soft-voting ensemble achieves a macro Area Under the Curve (AUC) of 95.92% and an overall accuracy of 80.24% on the held-out test set. To support qualitative interpretation, Gradientweighted Class Activation Mapping (Grad-CAM) is used to visualize model attention. In addition, a dedicated melanoma-screening analysis attains a Negative Predictive Value (NPV) of 97.1% at the selected operating threshold, highlighting the potential of the framework as a decision-support aid rather than a standalone diagnostic system.</span></p> <p class="MsoNormal"><strong><span>Keywords:</span></strong><span> Skin Lesion Classification, Deep Learning, CBAM Attention, Ensemble Learning, Grad-CAM, HAM10000, Clinical Triage</span></p> <p class="MsoNormal"><span> </span></p>
title Bridging the Black Box: An Interpretable Deep Learning Framework for Multi-Class Dermatological Screening
url https://doi.org/10.5281/zenodo.20259265