Mapping the Safety Boundaries of Medical AI: A Mechanistic Failure Analysis of Pneumonia Detection Ensembles under Severe Domain Shift
3rd International Conference on Information Technologies and Their Applications (ITTA 2026 Conference), Baku, Azerbaycan, 22 - 24 Nisan 2026, ss.1-15, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: Baku
- Basıldığı Ülke: Azerbaycan
- Sayfa Sayıları: ss.1-15
- Ankara Üniversitesi Adresli: Evet
Özet
Automated pneumonia detection faces critical barriers due to domain shift and AI brittleness, as deep learning models often struggle to generalize across disparate clinical environments and varied imaging techniques. This study introduces a triple-ensemble framework (VGG16, ResNet50V2, MobileNet V2) utilizing a robust 1,047-image validation partition and balanced loss weighting. While achieving 91.99% internal accuracy, the research distinguishes itself by pivoting to a failuremode analysis through independent external validation, where performance collapsed to 51.00%. A primary advantage of this work is the systematic exposure of diagnostic unreliability via Grad-CAM mapping, which identifies shortcut learning anchored to non-pathological hospital artifacts rather than anatomical signatures. Furthermore, the study provides empirical evidence of representational collapse; targeted fewshot adaptation yielded negligible recovery defining the definitive limits of shallow adaptation. By transitioning from performance maximization to mechanistic failure analysis, this research advances the literature with a reproducible framework for mapping safety boundaries in medical AI. These advantages provide a necessary roadmap for identifying boundary conditions of trust, advocating for deep-adaptation strategies in heterogeneous healthcare environments. It moves the discourse beyond accuracy-chasing toward a trust-centered methodology essential for deploying safety-critical medical systems.