MSc Thesis Defense: Onur Can, EXPLAINABILITY AND RELIABILITY IN MEDICAL AI SYSTEMS, Date & Time: 21 July, 2026 – 11:00 AM, Place: FENS L045

EXPLAINABILITY AND RELIABILITY IN MEDICAL AI SYSTEMS

Onur Can
Computer Science and Engineering, MSc Thesis, 2026

Thesis Jury

Prof. Berrin Yanıkoğlu (Thesis Advisor)

Prof. Erchan Aptoula

Prof. Mehmet Burçin Ünlü

Date & Time: July 21st, 2026 – 11:00 AM

Place: FENS L045

Keywords : Bone Fracture Detection, Deep Convolutional Networks, Ensemble

Learning, Explainable AI (XAI), Medical Decision Support Systems

Abstract

Bone fractures are among the most common presentations in emergency radiology, yet their detection from X-ray images remains limited by interpreter variability, severe class imbalance, and the difficulty of trusting automated predictions. This thesis develops an evidence-based fracture-detection system on the FracAtlas dataset (4,083 radiographs), addressing the two gaps that separate a high-accuracy model from clinical use: explainability and reliability. A systematic modeling progression is pursued with Convolutional Neural Network (CNN) classification baseline, YOLOv8 localization that supplies spatial evidence a clinician can verify, integration into FracAssist (a deployed, reproducible interactive system), and a structured ablation of twenty-five ensemble methods organized under a five-family taxonomy. The localization detector reached a mAP@0.5 score of 65.14%, exceeding the published FracAtlas baseline by +8.94%. Furthermore, no ensemble strategy surpassed a validation AUC ceiling score of 90.60%, attained by the integrated combiner named Gated Ensemble Logic (GEL) with a validation F1 score of 73.94%, which resolves calibration tension and is presented as the validated endpoint of that comparison. Therefore, this thesis addresses two primary pain points by addressing the lack of explainability through verifiable spatial tracking and resolving the issue of model reliability through a transparent evaluation of multiple ensemble techniques that highlight both their strengths and actual limitations. To ensure transparency, the complete pipeline is released publicly for full reproducibility.