EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Deep Learning for Small Bowel Lesion Classification: A Class-Aware Baseline Addressing Long-Tailed Clinical Distributions

Zihan Zhang

IEEE Access · 2026 · 14 · 97331-97340

DOI: 10.1109/access.2026.3706419 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Zihan Zhang. Die Arbeit nennt 33 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The differential diagnosis of small bowel diseases poses a significant clinical challenge, particularly in distinguishing Crohn’s disease (CD) from intestinal tuberculosis (ITB) due to their highly overlapping endoscopic morphologies. Moreover, the inherent long-tailed distribution of clinical datasets—where rare lesions are substantially underrepresented—severely impedes the generalization capability of conventional deep learning models. In this paper, we propose a class-aware deep learning baseline built on the ConvNeXt V2 architecture. A total of 1,790 white-light endoscopic images spanning five diagnostic categories (CD, CMUSE, ITB, NSU, and PSIL) were retrospectively collected. To address the extreme class imbalance (e.g., a 14:1 ratio between CD and PSIL), a Weighted Random Sampling (WRS) strategy was integrated into the training pipeline. Due to retrospective anonymization constraints, the model was evaluated using an image-level 5-fold cross-validation protocol. The proposed framework achieved an overall accuracy of 88.72% $\pm ~1.19$ % and a Macro F1-score of $0.8353~\pm ~0.0270$ . Ablation studies comparing sampling strategies (WRS vs. uniform), loss functions (cross-entropy, focal loss, class-balanced loss), and backbone architectures (ConvNeXt V2, ResNet-50, EfficientNet-B4) confirmed that the ConvNeXt V2 + WRS configuration achieved the highest overall performance. However, differentiating ITB from CD remains a fundamental performance bottleneck, with an ITB recall of 64.91%, reflecting a persistent challenge rooted in the morphological similarity between these two diseases. We acknowledge that the image-level validation protocol, in the absence of patient identifiers, introduces a risk of intra-patient data leakage. By explicitly addressing long-tailed distributions and comprehensively benchmarking alternative strategies, the proposed framework establishes a transparent image-level baseline for small bowel lesion classification, empirically highlighting the performance ceiling of single-modality visual models and underscoring the necessity for future patient-level, multi-modal diagnostic approaches.

Zitieren

APA

Zhang, Z. (2026). Deep Learning for Small Bowel Lesion Classification: A Class-Aware Baseline Addressing Long-Tailed Clinical Distributions. IEEE Access, 14, 97331-97340. https://doi.org/10.1109/access.2026.3706419

Vancouver

Zhang Z. Deep Learning for Small Bowel Lesion Classification: A Class-Aware Baseline Addressing Long-Tailed Clinical Distributions. IEEE Access. 2026;14:97331-97340. doi:10.1109/access.2026.3706419

BibTeX
@article{Zhang2026,
  author = {Zhang, Zihan},
  title = {Deep Learning for Small Bowel Lesion Classification: A Class-Aware Baseline Addressing Long-Tailed Clinical Distributions},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {97331--97340},
  doi = {10.1109/access.2026.3706419}
}
Kostenlos starten →