Uncertainty-aware bayesian convolutional neural network for breslow thickness classification from dermoscopic images

Name: Agam, S.T., Regan

Organization : Gadjah Mada University. Department of Electrical Engineering & Information Technology

Name: Wicaksono, Ridwan

Organization : Gadjah Mada University. Department of Electrical Engineering & Information Technology

Email : ridwan.wicaksono@mail.ugm.ac.th

Name: Nugroho, Hanung Adi

Organization : Gadjah Mada University. Department of Electrical Engineering & Information Technology

Email : adinugroho@ugm.ac.id

keyword: Convolutional neural networks

LCSH: Melanoma -- Diagnosis

LCSH: Skin -- Cancer -- Diagnosis

LCSH: Diagnostic imaging -- Digital techniques

LCSH: Neural networks (Computer science)

LCSH: Bayesian statistical decision theory

Abstract: Breslow thickness is the most critical prognostic determinant in melanoma staging. Estimating this continuous tumor invasion depth from two-dimensional dermoscopic images is an inherently ambiguous inverse problem. Consequently, standard deterministic neural networks often yield overconfident, miscalibrated predictions, limiting their utility in clinical decision-making. Furthermore, leveraging unlabeled data through semi-supervised learning (SSL) frequently degrades performance due to pseudo-label noise amplification near the decision boundary. To address these challenges, this paper proposes an uncertainty-aware bayesian convolutional neural network (UC-BCNN) to classify breslow thickness at the critical 0.8 mm threshold. We compare three experimental settings (A0: supervised, A1: standard semi-supervised, and A2: UC-driven bayesian semi-supervised) using a ResNet-50 backbone. Evaluations on a composite dataset (868 clinical VRUH images, 545 labeled ISIC images, and 2,000 unlabeled ISIC images) using strict 5-fold cross-validation reveal that the naive SSL baseline (A1) fails to yield consistent gains due to label noise. Conversely, the proposed UC-BCNN (A2) employs Monte Carlo-driven uncertainty filtering to reject highly ambiguous pseudo-labels. The A2 model achieves a balanced accuracy of 0.947, an F1-score of 0.954, and a Cohen’s Kappa of 0.892. These findings demonstrate that integrating epistemic uncertainty quantification aligns predictive models with clinical risk minimization, effectively reducing false negatives and offering a robust foundation for AI-assisted staging.

King Mongkut's University of Technology North Bangkok. Central Library

Address: BANGKOK

Email: library@kmutnb.ac.th

Created: 2026

Modified: 2026-08-17

Issued: 2026-08-17

บทความ/Article

application/pdf

BibliograpyCitation : In ECTI Association and Burapha University. Faculty of Engineering. 2026 23rd International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON 2026) (pp.988-993). Piscataway, NJ : IEEE

eng

©copyrights King Mongkut's University of Technology North Bangkok

ลำดับที่.	ชื่อแฟ้มข้อมูล	ขนาดแฟ้มข้อมูล	จำนวนเข้าถึง	วัน-เวลาเข้าถึงล่าสุด
1	ECTI-CON 2026pp.988-993.pdf	1.54 MB

ใช้เวลา

0.043727 วินาที