What question did this study set out to answer?

To systematically review CNN models trained on RGB photographs for tasks related to burn segmentation and classification.

February 22, 2026Open Access

Convolutional neural networks for burn segmentation and classification tasks using RGB photographs: a five-year systematic review

Key Points

To systematically review CNN models trained on RGB photographs for tasks related to burn segmentation and classification.
Conducted a systematic search in PubMed, Medline, Embase, and Cochrane Library from January 2020 to April 2025.
Included studies applying CNN models to RGB images for burn area segmentation, burn depth classification, or burn depth segmentation.
Assessed risk of bias using PROBAST + AI and extracted data on model architecture and performance metrics.
Six studies reported burn area segmentation accuracy above 90%.
Four studies achieved burn depth classification accuracy above 80%, with the best models using feature enhancement strategies reaching up to 98%.
Low-quality data and annotation inconsistencies negatively impacted burn depth segmentation performance.

Abstract

• Focuses exclusively on CNN models trained on RGB photographs. • ResNet-101 combined with ASPP provided a strong baseline in burn area segmentation. • Feature enhancement techniques improved model performance in burn depth classification. • Dataset quality and annotation consistency influenced model performance in burn depth segmentation. Burn wound assessment remains complex and often inaccurate, with visual evaluations by non-specialists achieving < 50% accuracy. Although Laser Doppler Imaging (LDI) offers accuracy rates of up to 99%, its high cost and distrust by clinicians limit widespread adoption. Convolutional neural networks (CNNs) offer a promising alternative: when trained on RGB images, they enhance accessibility, reduce assessment time, and improve consistency among clinicians. This review systematically evaluates CNN-based approaches for burn area segmentation (BAS), burn depth classification (BDC), and burn depth segmentation (BDS) using RGB photographs. A systematic search of PubMed, Medline, Embase, and Cochrane Library (January 2020–April 2025) was conducted on 1st April 2025. Studies applying CNNs to RGB images for one or more of the prediction tasks (BAS, BDC, or BDS) were included. Risk of bias was assessed using PROBAST + AI. Data on model architecture, datasets, and performance metrics were extracted and synthesized narratively. A total of 14 studies were included. Among BAS tasks, six reported accuracy above 90%. The combination of ResNet-101 and Atrous Spatial Pyramid Pooling (ASPP) provides a strong and stable baseline across studies. In BDC tasks, four reported an accuracy above 80%. The top two best-performing models employed feature enhancement strategies to achieve accuracy up to 98%. In BDS tasks, low-quality data and inconsistent annotation were observed to negatively affect model performance. This review underscores the reliability and potential architecture components of CNN models trained on RGB photographs, supporting the future integration of such models into mobile phone–based applications.

Convolutional neural networks for burn segmentation and classification tasks using RGB photographs: a five-year systematic review

Key Points

Abstract

Cite This Study