AI 의료기기 소프트웨어 평가 방법 IT융합대학원 한 상 택 지 도 교 수 정 기 현 본 논문은 AI 의료기기 소프트웨어의 평가 방법에 관한 연구이다. AI 기술 의 발전으로 인해 의료 분야에서 AI 의료기기 소프트웨어의 활용이 더욱 활발 해지면서, 소프트웨어의 안전성과 유효성을 검증할 필요성이 대두되고 있다. 이 에 본 논문에서는 기존 AI 의료기기 소프트웨어의 평가 방법들의 단점을 보완 하는 방법을 제안한다. 제안하는 방법에서는 소프트웨어의 학습 및 예측 능력 을 검증하기 위한 추가적인 측정 지표를 제안한다. 제안하는 측정 지표로는 적대적 예제 훈련, 교정, 불확실성의 정량화 방법 으로 적대적 예제 훈련의 경우 성능 향상, 적대적 저항성 평가가 가능하다. 또 한, 모델 성능이 기준보다 개선되었는지와 성능 저하가 없었는지를 확인한다. 교정에서는 모델의 예측 확률이 제조자가 제시한 기준보다 높은지 확인하며 불 확실성의 정량화에서는 모델의 불확실성 예측 확률이 기준값 이상인지 확인 가 능하다. 제안하는 지표의 유효성과 현장 적용 가능성 확인을 위하여 사례 검증 을 시행한다.|This thesis examines the evaluation methods for AI medical device software. With the advancement of AI technology, the utilization of AI medical device software in the medical field has become more prominent, necessitating verification of the safety and efficacy of such software. In the thesis, a method is proposed to address the shortcomings of existing evaluation methods for AI medical device software. The proposed method suggests additional metrics to assess the software's learning and prediction capabilities._x000D_
<br>_x000D_
<br>As part of the proposed metrics, methods for adversarial example training, calibration, and quantification of uncertainty are introduced. Adversarial example training allows for performance enhancement and evaluation of adversarial robustness. Additionally, it examines whether the model's performance has improved compared to benchmarks without performance degradation. Calibration checks if the model's prediction probabilities exceed those specified by the manufacturer, while uncertainty quantification assesses if the model's uncertainty prediction probabilities meet predefined criteria. Case validation is conducted to verify the effectiveness and practicality of the proposed metrics and their applicability in real-world scenarios.
Alternative Abstract
This thesis examines the evaluation methods for AI medical device software. With the advancement of AI technology, the utilization of AI medical device software in the medical field has become more prominent, necessitating verification of the safety and efficacy of such software. In the thesis, a method is proposed to address the shortcomings of existing evaluation methods for AI medical device software. The proposed method suggests additional metrics to assess the software's learning and prediction capabilities._x000D_
<br>_x000D_
<br>As part of the proposed metrics, methods for adversarial example training, calibration, and quantification of uncertainty are introduced. Adversarial example training allows for performance enhancement and evaluation of adversarial robustness. Additionally, it examines whether the model's performance has improved compared to benchmarks without performance degradation. Calibration checks if the model's prediction probabilities exceed those specified by the manufacturer, while uncertainty quantification assesses if the model's uncertainty prediction probabilities meet predefined criteria. Case validation is conducted to verify the effectiveness and practicality of the proposed metrics and their applicability in real-world scenarios.
