EVALUATING PREDICTIVE CERTAINTY IN AI MODELS FOR ACCURATE BRAIN TUMOR DETECTION

Brain tumors are highly dangerous and often life-threatening, significantly affecting patients’ overall health and quality of life. This research explores prediction certainty, an underrepresented area, as existing research focuses on accuracy. This study highlights establishing a correlation between a model’s loss value and greater certainty in predictions. Along with conventional performance metrics such as precision and recall, this work emphasizes the critical role of loss value. To assess the reliability and effectiveness of artificial intelligence models, including CNN, ResNet-50, XceptionNet, and a proposed model (integrating advanced layers), were tested. The study prioritized loss values for accurate detection of tumor cases, minimizing false negatives. The models are effective for real-time tumor detection due to their low loss values and efficient runtimes. Experimental results showed the following metrics on testing: CNN achieved a loss of 0.35 and 68.60% accuracy; ResNet-50 achieved improved performance with a loss of 0.17; and the proposed model achieved 90% accuracy with superior recall and runtime. The study concludes that while accuracy is important, the certainty in predictions plays a significant role in reliable tumor diagnosis. Given the global shortage of specialized medical professionals, the proposed approach addresses this gap by providing timely and accurate cancer detection tools, contributing effectively to healthcare systems and medical education, enabling future AI applications to be used effectively in clinical practice.