Task-specific regularization loss towards model calibration for reliable lung cancer detection
This work addresses the need for reliable automated lung cancer detection to assist radiologists in high-pressure environments like India, though it is incremental as it builds on existing calibration methods.
The paper tackled the problem of deep neural networks making overconfident mistakes in lung cancer detection from CT scans, proposing a task-specific regularization loss combined with MDCA loss and post-hoc calibration to achieve a 5.98% improvement in Expected Calibration Error and a 17.9% improvement in Maximum Calibration Error compared to the best-performing state-of-the-art algorithm.
Lung cancer is one of the significant causes of cancer-related deaths globally. Early detection and treatment improve the chances of survival. Traditionally CT scans have been used to extract the most significant lung infection information and diagnose cancer. This process is carried out manually by an expert radiologist. The imbalance in the radiologists-to-population ratio in a country like India implies significant work pressure on them and thus raises the need to automate a few of their responsibilities. The tendency of modern-day Deep Neural networks to make overconfident mistakes limit their usage to detect cancer. In this paper, we propose a new task-specific loss function to calibrate the neural network to reduce the risk of overconfident mistakes. We use the state-of-the-art Multi-class Difference in Confidence and Accuracy (MDCA) loss in conjunction with the proposed task-specific loss function to achieve the same. We also integrate post-hoc calibration by performing temperature scaling on top of the train-time calibrated model. We demonstrate 5.98% improvement in the Expected Calibration Error (ECE) and a 17.9% improvement in Maximum Calibration Error (MCE) as compared to the best-performing SOTA algorithm.