EUVIMEDEUVIMED entdecken
Crossref

Retrieval Familiarity for Selective Referral in Dermoscopic Classification: An External Evaluation Under Distribution Shift

Pronab Kumar Mondal, Subrata Kumar Das, Sumon Kumar Debnath, Indrani Mandal, Md. Selim Al Mamun

International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)

DOI: 10.14569/ijacsa.2026.01710108 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Pronab Kumar Mondal und 4 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This study audits retrieval familiarity as a post-hoc accept–refer signal for dermoscopic classification under dataset shift. Mean Top-5 familiarity was defined as the average cosine similarity between a query and its five nearest cases in a lesion-disjoint HAM10000 reference bank. Retrieval depth and the operating threshold were selected on HAM10000 validation data and then fixed. Evaluation used 14,885 ISIC 2019 images without matching development identifiers and 963 eligible Derm7pt images. The base classifier achieved 53.83% accuracy (macro-F1 0.4012) on ISIC 2019 and 64.38% accuracy (macro-F1 0.3926) on Derm7pt. Mean Top-5 familiarity obtained an area under the risk–coverage curve (AURC) of 0.3498 on ISIC 2019 and 0.2514 on Derm7pt, ranking fourth and ninth, respectively. A Combined Random Forest error score, constructed using consistent feature definitions across all cohorts, achieved lower AURCs of 0.3245 and 0.2096; the paired differences from Mean Top-5 were −0.0253 (95% CI −0.0324 to −0.0184; Holm-adjusted p = 0.013) and−0.0419 (−0.0670 to −0.0162; Holm-adjusted p = 0.022). A threshold designed for 50% validation coverage produced only 21.95% and 39.36% coverage for Mean Top-5 and 19.73% and 37.59% for the combined score. At 50% ISIC coverage, the familiarity rule accepted 1,034 of 2,153 misclassified melanomas. Predicted-class and cost-weighted analyses also showed that aggregate ranking did not establish clinical safety. The results support retrieval familiarity as an inspectable but weak signal whose value improved when combined with classifier information. They also show that score definitions, feature pipelines, thresholds, and class-level behaviour require careful external auditing before clinical use.

Zitieren

APA

Mondal, P. K., Das, S. K., Debnath, S. K., Mandal, I., & Mamun, M. S. A. (2026). Retrieval Familiarity for Selective Referral in Dermoscopic Classification: An External Evaluation Under Distribution Shift. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.01710108

Vancouver

Mondal PK, Das SK, Debnath SK, Mandal I, Mamun MSA. Retrieval Familiarity for Selective Referral in Dermoscopic Classification: An External Evaluation Under Distribution Shift. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.01710108

BibTeX
@article{Mondal2026,
  author = {Mondal, Pronab Kumar and Das, Subrata Kumar and Debnath, Sumon Kumar and Mandal, Indrani and Mamun, Md. Selim Al},
  title = {Retrieval Familiarity for Selective Referral in Dermoscopic Classification: An External Evaluation Under Distribution Shift},
  journal = {International Journal of Advanced Computer Science and Applications},
  year = {2026},
  volume = {17},
  number = {10},
  doi = {10.14569/ijacsa.2026.01710108}
}
Kostenlos starten →