TWD-Det: uncertainty-aware three-way decision framework for robust medical object detection
TL;DR
By strictly lowering miss rates, TWD-Det provides an optimal interface for Human-in-the-loop (HITL) medical AI software, demonstrating tremendous engineering value.
Abstract
In medical image analysis, lesions often exhibit blurred boundaries, causing traditional binary object detectors to produce clinically unacceptable missed detections and false positives. To address this, we propose TWD-Det, a robust, uncertainty-aware framework based on Three-Way Decision (TWD) theory. Using an uncertainty quantification mechanism and a dual-threshold strategy (α, β), TWD-Det reconstructs the conventional binary output into a ternary space: acceptance, rejection, and a boundary buffer zone. High-uncertainty predictions falling into the boundary region are routed to a lightweight Refine Module, which performs secondary fine-grained reasoning using broader contextual features to achieve a "deferred decision." Evaluated as a plug-and-play component on the LUNA16 dataset using YOLOv8, TWD-Det significantly reduced the Miss Rate (False Negative Rate, FNR) from 14.8% to 8.3% and increased Recall to 91.7% with minimal overhead (+0.12M). Breaking the hard-classification paradigm, it naturally replicates clinical cognitive logic. By strictly lowering miss rates, TWD-Det provides an optimal interface for Human-in-the-loop (HITL) medical AI software, demonstrating tremendous engineering value.