Expert-level diagnosis of pediatric posterior fossa tumors via consistency calibration

Accurate diagnosis of pediatric posterior fossa tumors (PFTs) is critical for saving lives; however, the limited number of specialists makes accurate diagnostics scarce. To make the diagnosis of PFTs accurate, automatic, and noninvasive, scholars have proposed employing deep neural networks (DNNs) to predict tumor types using magnetic resonance imaging data. Advanced methods primarily focus on fine-tuning DNNs pre-trained on large-scale datasets of natural images, e.g., ImageNet. However, the existing methods overlook the priors of human experts. Human experts typically recheck whether images predicted as a particular class are similar to those predicted as the same class to ensure prediction consistency. Therefore, the predicted results of an intelligent system should be consistent. Inspired by the rechecking process, we propose a novel learning paradigm called Consistency calibration (Coca). Within the Coca framework, the output predicted by DNNs is guided by two objective functions: (i) the task-specific objective of making the predicted results the same as the groundtruth, and (ii) an auxiliary objective of rechecking the prediction consistency. Coca is developed by defining the inconsistency for each sample by inconsistent risks: the auxiliary risk is small (large), but the task-specific risk is large (small). Building on the inconsistency definition, Coca identifies inconsistencies for each sample using an adversarial attack. Subsequently, these inconsistencies are leveraged to tune DNNs in an adversarial training manner for consistency calibration. To verify the efficacy of Coca, we conduct comprehensive experiments using a large-scale PBT dataset, and the results show that Coca significantly outperforms state-of-the-art methods. Moreover, Coca has improved performance over human experts as demonstrated by expert-level diagnostic performance in real-world PBT scenarios for the first time.